Set up and run the platform on your site

How the platform works

The platform is AI video analytics for the CCTV you already have. It runs on one server at your site, with no cloud, and watches every camera for missing protective equipment (PPE), fire and smoke, people in restricted zones, risky moves near forklifts, people and vehicle counts, number plates, visitors and more. You switch the checks on per camera. Each alert carries a picture and a clip, reaches the right people, and can become an incident to assign, resolve and hand over at the end of the shift. Video, faces and site data stay on the site.

The steps below follow the PPE check, the platform's core check. When someone lacks a required item for long enough, the platform records an event with evidence and tells the people who are watching. The other checks follow the same path from camera to evidence to review. People decide: an event is evidence for a person to review, and a new model never goes live until a person activates it.

1. Cameras send video. The platform reads each camera's stream and records it to the server's disk.

2. People are detected and tracked. Each person keeps one id across frames and cameras (chapter 13).

3. PPE is checked against the zone's rules. The zone decides which items a person needs and how serious a miss is (chapter 7).

4. A violation is confirmed over a few seconds. An item must stay missing for about 8 seconds by default, so one blurry frame raises no alarm (chapter 10).

5. An event is created with evidence. The platform saves two snapshots and a short clip, and the event appears in the bell and in Events.

6. A person reviews it. A reviewer marks the event as a real violation or not a violation (chapter 11).

7. The model learns from reviewed frames. Approved frames train a new candidate model, and a person decides whether to activate it (chapter 16).

The parts, for the person who installs it

Everything you would want to keep lives in Postgres and in one data folder on the SSD (/data/ppe in the Docker install). Settings such as thresholds, rules and zones are data edited in the screens. They reach the vision worker within 15 seconds and need no restart.

How accurate is it?

Treat the first weeks as a tuning period. Review events, mark the wrong ones as not a violation, and let the review queue fill. Camera placement matters too: people who appear small, far away or seen from high above are the hardest to judge.

Where to start

• AdministratorsChapters 4 to 8, then 16 and 17.
• The person who installs itChapters 2 and 3, then 17 and 18.
• SupervisorsChapters 5 to 7 and 9 to 15.
• ReviewersChapters 9, 11, 12 and 13, and the Review queue and Accuracy checks in chapter 16.

Before you install

Check the server, the cameras, the network and the browser before you start.

The server

Choose a server size

Recordings take about 43 GB per camera per day at 4 Mbit/s, so size the recordings drive for the days you keep. The tiers name a use, not a camera count: Run a load test on the System page says how many cameras fit on your server (chapter 17).

Example: 4 cameras kept for 4 weeks

Estimates. They leave the default 50 GB of free space and assume continuous recording. Recording the sub stream or a lower bitrate stretches the days; the Space saving mode on the Storage page does that automatically before the disk fills (chapter 17).

The cameras

Each camera needs an RTSP address and its login, reachable from the server. Use the camera's low-resolution sub stream for detection and its main stream for recording. The platform decodes the sub stream, and it records the main stream without re-encoding it. A camera that must find small or far people can read its main stream for detection instead (chapter 6). A Hikvision camera, for example, offers both:

Main stream   rtsp://user:password@192.0.2.10:554/Streaming/Channels/101
Sub stream    rtsp://user:password@192.0.2.10:554/Streaming/Channels/102

Before you add a camera, test the address from the server. If this command ends without an error, the address works:

ffmpeg -rtsp_transport tcp -i URL -t 5 -f null -

The platform does not judge people who are shorter than 80 pixels in the picture (a setting you can change). Aim and zoom each camera so workers appear large enough to show a helmet and a vest.

Network ports

With a DNS name in PUBLIC_HOST that points at the static address, Caddy obtains a Let's Encrypt certificate. With an IP address or localhost it uses its own certificate authority, and you import caddy_data/pki/authorities/local/root.crt on each viewing computer. Viewers outside the LAN also need port 8555 reachable from outside, and the server's public address set as webrtc.candidates in infra/go2rtc.yaml.

The browser

The platform runs in the browser, so nothing is installed on viewing computers. Use Chrome 111, Edge 111, Firefox 128 or Safari 16.4 or later, on a computer, an iPhone or an iPad. Live video on an iPhone through the server needs iOS 17.1 or later. Three features ask for something extra. Live video on the wall needs WebRTC and port 8555. Kiosk needs the browser's full-screen mode. Browser notifications need you to allow notifications for the site.

Install

The platform runs on Ubuntu, Windows or macOS. Choose the route for the server you have. Other devices on the local network open the platform at the server's addresses or its .local name, found over Bonjour.

• Windows
• macOS
• Linux
• Docker

Ubuntu: Docker Compose with an NVIDIA GPU

Plan one to two hours, most of it downloads. The steps below summarise docs/INSTALL-LINUX-GPU.md, which gives every command in full. First, know the five secrets you will create in step 4. Compose refuses to start while any of them is empty, and in production the API refuses placeholders and short values.

1. Install the NVIDIA driver, version 580 or newer. Reboot, then check that nvidia-smi prints your GPU and a driver version of 580 or higher.

2. Install Docker Engine, the Compose plugin and the NVIDIA container toolkit. Then run this test. It must print the same table as the host. If it fails, nothing below can see the GPU, so fix it first.

   docker run --rm --gpus all ubuntu:24.04 nvidia-smi

3. Get the code and create the data folders.

   sudo apt-get install -y git jq postgresql-client
   sudo install -d -o "$USER" /opt/ppe /data/ppe /var/backups/ppe
   git clone <the repository url> /opt/ppe

4. Create the settings file and fill in the five secrets. Make each value with the command in the table. Also set DATA_DIR. For a LAN site served on port 3000, set EXTRA_WEB_ORIGINS=http://<server address>:3000 so the API accepts sign-ins from that page. For a public name with HTTPS, set PUBLIC_HOST and ACME_EMAIL instead.

   cd /opt/ppe
   cp infra/.env.compose.example infra/.env && chmod 600 infra/.env

5. Check the file before go-live. The check flags weak or missing secrets and prints no values.

   scripts/prod-check.sh infra/.env

6. Start the platform. The first build downloads PyTorch and the CUDA libraries, several GB, and takes 10 to 30 minutes. A new site downloads the pinned D-FINE small person detector by itself. To use your own person model instead, put it in /data/ppe/models/base/, or install one from the Models page later.

   mkdir -p /data/ppe/models/base
   scripts/compose.sh up -d --build postgres redis go2rtc api vision trainer web-lan

   For HTTPS with Caddy, name web caddy instead of web-lan. Running scripts/compose.sh up -d --build with no names starts every service in the file. If ufw is on, open the ports from the table in chapter 2.

7. Verify. Every service should show running and Postgres healthy. The second command answers with {"ok":true,...}.

   scripts/compose.sh ps
   curl -fsS http://localhost:4000/api/health

   Then sign in (chapter 4) and open System. Vision shows a recent heartbeat and the provider torch-cuda. If it shows torch-cpu, the driver is older than 580, the GPU is Pascal or older, or the container toolkit is missing. The vision log says which: scripts/compose.sh logs vision.

8. Turn on the nightly backup. It runs at 02:30 UTC by default (chapter 17).

   scripts/compose.sh --profile backup up -d backup

If Compose stops with bind source path does not exist: /etc/machine-id, the server has no machine id. Run sudo systemd-machine-id-setup and start again.

Windows and macOS: background services from the download

The download for Windows or for a Mac with Apple Silicon installs the platform as background services that start with the computer. It has its own built-in database and key-value store, so nothing else is installed. On a Mac, detection uses the Mac's own machine-learning chip. Unpack the download, open an administrator PowerShell (Windows) or Terminal (macOS), go to the unpacked folder and run .\bin\platformctl.exe install --from (Windows) or sudo ./bin/platformctl install --from (macOS) followed by the path of the downloaded .zip. It checks the package, asks before it installs, and ends by showing the first administrator password once. On Windows, other computers can open the platform only after an administrator allows ports 3000 and 4000 in Windows Firewall; the README.txt has the one line for it.

Each system has an install file as well as its zip: the Windows setup (.exe), the macOS installer (.pkg) and the Ubuntu and Debian package (.deb). Open it for a new computer, or for a site installed from the zip. The Windows setup needs no Unblock step. These downloads are not signed. The first time, Windows shows "Windows protected your PC": choose More info, then Run anyway. If there is no Run anyway, Smart App Control is on and blocks unsigned programs. On a Mac, remove the download mark with xattr -dr com.apple.quarantine on the unpacked folder, or choose Open Anyway in System Settings, Privacy & Security. The README.txt in the download has each step. To move from the earlier one-click install, add --adopt-r0: it keeps the database and passwords. Later versions install from Settings, Updates, or with platformctl update --from, and platformctl rollback goes back for 7 days.

Windows: Run Me.bat

First unpack the download into a short folder outside OneDrive, for example C:\ppe: Downloads, Desktop and Documents are often inside OneDrive, and a deep folder makes Windows refuse the longest file names. The installer checks this in its first minutes and says the exact move if the folder is wrong. Then double-click Run Me.bat. It runs scripts/dev.ps1, which uses winget (or Chocolatey where winget is missing) to install bun, Node.js, uv, ffmpeg, PostgreSQL 17 and a Redis-compatible server, creates the settings file and the ppe database, installs the dependencies, and starts the API, the web app, the vision worker and the trainer. Logs go to data\logs. The first start compiles the web app and takes a few minutes. If a step fails, for example because a service needs Administrator rights once or winget is missing, the window stays open with the error. Fix it and run the file again. The installer generates the administrator password, shows it once in the window and keeps it in the settings file (the ADMIN_PASSWORD line).

macOS: Run Me.command

Double-click Run Me.command. A downloaded copy is blocked the first time: right-click it and choose Open, or run xattr -d com.apple.quarantine "Run Me.command" in Terminal. The script checks for Homebrew, bun, uv, ffmpeg, PostgreSQL 17 and Redis, offers to install what is missing, starts everything and opens the browser when the API is healthy. bash "Run Me.command" --check only reports what is missing.

Ubuntu desktop or VM: Run Me (Linux).desktop

Right-click Run Me (Linux).desktop and choose Allow Launching (once), then double-click it. It asks for your password once, installs ffmpeg, PostgreSQL, Redis, bun and uv, creates the settings file and the ppe database, and starts the platform. The first start downloads several gigabytes of packages and can take 10 to 20 minutes; later starts take seconds. It runs on the processor, so no GPU is needed.

Ubuntu: system services without Docker

Clone to /opt/ppe, install the dependencies, build the web app, copy .env.example to /opt/ppe/.env, install the units from infra/systemd and enable ppe-api, ppe-web, ppe-vision and ppe-trainer. Set STORAGE_ROOT to an absolute path, because the units start in different folders. Section 9 of docs/RUNBOOK.md has the details.

Updates

• Upload a package in Settings. An administrator opens Settings → Updates and installs a new version from its package file (chapter 17, Backups and upgrades).
• Re-run a newer installer. It upgrades the platform in place and keeps your data and settings. It also keeps the previous version beside the new one, so you can go back, and removes only older versions. On a computer that already had the platform's database from an earlier setup, the installer takes that database over with everything in it; running it again repairs a computer where every page said it could not reach the server.
• Roll back in Settings → Updates. The earlier version is kept for 7 days. If a security update has set a lowest version, the Roll back button is locked and the page says why.
• Online updates. An administrator opens Settings → Updates and turns on Check automatically to look in an update folder every few hours, or presses Check now. A newer version for this server is downloaded, and its signature is checked before Install is offered. Install asks for the administrator's password. With Install automatically on, a checked update installs by itself in the window the administrator chooses, when no training, analysis or export is running. Both are off until an administrator turns them on.

Release packages ship no source code.

More guides in the repository

• docs/INSTALL-LINUX-GPU.md: a new GPU server from scratch.
• docs/RUNBOOK.md: operations, settings, models, storage and troubleshooting.
• docs/DEPLOYMENT.md: remote access through a Cloudflare Tunnel, and the security checklist.
• docs/BACKUP.md and docs/UPGRADE.md: backups, restores, upgrades and rollback.

First sign-in and first hour

This chapter takes a fresh install to a camera that checks PPE. Plan about an hour.

1. Sign in. Open http://<server address>:3000 on a LAN install, or https:// and your PUBLIC_HOST name when Caddy is in front. Sign in with the administrator account. Tick Keep me signed in on your own computer so the platform does not ask again for the number of days it names; leave it unticked on a shared one. The password field warns you when Caps Lock is on, and a failed sign-in says what went wrong.

2. Change the administrator password. Open the account menu at the bottom of the sidebar and choose Change password. Enter the current password and a new one of at least 10 characters twice, then choose Change password. Your other sessions are signed out.

3. Set your site name. Open Settings → General & branding. Enter a Site name (shown in the top bar, hidden when empty) and choose Save changes. The Product name, Accent colour and Logo come from your licence. If the licence includes Own branding, an administrator can change them here (the logo is a PNG or SVG of up to 64 KiB) and Reset branding returns to the licence's own. Otherwise they show read-only, and you ask whoever supplied the licence. A Brand preview shows the sign-in page.

4. Create users. Open Users and choose Invite. Enter a name, an email and a role (chapter 5), choose Create invitation and send the person the link it shows, so they choose their own password. Add user sets a password for them instead. Give each person the lowest role that lets them do their job.

5. Activate a PPE model. Open Models. In the Model library, Download a ready-made PPE detector, then choose Activate. The first activation needs an override because no golden-set score exists yet: enter a reason such as "first install" (chapter 16). Without an active PPE model, the platform tracks people but judges nothing, and the System page shows the no_ppe_model warning.

6. Add your first camera. Open Cameras and choose Add camera. Enter a name, the main stream address and the camera's login. Add the sub stream address if the camera has one. Keep Checks at every check, Checks per second at 5 and Record on, then choose Add camera. The Status column changes to Online within moments. The platform creates a zone called "Whole frame" for the camera.

7. Draw a zone if parts of the picture need different rules. Open the camera, choose the Zones tab and choose Add zone (chapter 6). Skip this step if the same rules apply everywhere.

8. Tick the required PPE. Open Rules. In the camera's row, click the cells for the items that zone requires, such as Safety helmet and Safety vest. Use the arrow in a cell to set the severity (chapter 7).

9. Check the Monitor. Open Monitor. You should see the live picture with a tag on each person. To test without waiting for a real miss, step in front of the camera without a required item for several seconds, or add a stored video as a camera (chapter 6). An event appears in Events a few seconds after the item has been missing.

My account: two-step sign-in and signed-in devices

Choose My account in the account menu. Two-step sign-in asks for a 6-digit code from an authenticator app on your phone after your password. Choose Set up two-step sign-in, enter your password, scan the QR code (or type the key under it), enter the code your phone shows and choose Turn on. The platform then shows ten recovery codes: copy them or download them as a text file, keep them somewhere safe and tick the box before you choose Done. Each recovery code signs you in once if the phone is lost. Make new recovery codes and Turn off ask for your password and a code. An administrator can require two-step sign-in for administrators in Settings → Sign-in & access, and the platform then asks them to set it up before anything else. Wall-screen accounts sign in without a code.

Signed-in devices on the same page lists every browser signed in as you, with its address and device, and lets you sign one out or all the others. A one-time link for a reset or an invitation can be shown as a QR code, so it reaches a phone without typing.

If someone loses their phone and their recovery codes, an administrator opens Users and chooses Turn off two-step sign-in on their row. Every device of theirs is signed out, and they sign in with their password only until they set it up again.

If you cannot sign in

• Forgotten password. Choose Forgot password? on the sign-in page and enter your email. The page gives the same answer whether or not the address has an account, and your administrators are told. One of them hands you a one-time link: it lets you choose a new password and signs you out everywhere else. An expired or used link says so and lets you ask again. It needs no email server.
• No account yet. When the site allows it, choose Request access and enter your name, email and a password, with an optional note on who you are. A strength meter helps you pick a good password, and each problem shows under its own field. You see nothing until an administrator approves the request and picks your role. Until then, signing in shows Waiting for approval with the date you asked; the page checks every few seconds and opens the platform the moment you are approved.
• A lost phone. At the code step choose Use a recovery code instead and enter one of the ten codes you saved. If you have none left, ask an administrator to turn off your two-step sign-in.
• A temporary password. If an administrator gives you one, sign in with it. The platform then asks you to Choose a new password before anything else.

Finding your way

The sidebar groups pages by the job they do. You see only the pages your role can use.

The menu and the command palette show only the apps that are switched on (Apps, in chapter 17). The AI pages (Rules, Events, Models) leave the menu when no AI app is on. A page of an app that is switched off says This app is switched off: administrators get an Open Apps button, and everyone else is asked to ask an administrator.

Press ⌘B on a Mac or Ctrl B elsewhere to collapse the sidebar to icons. The Events entry carries a count of open alerts, and the Datasets entry carries the number of samples waiting for review.

Near the bottom, Help opens Documents (this manual, the quick reference and the other documents) in a new tab. It opens the copy your installation serves at /docs, so it works offline. A development build also shows a Live demo, a clickable preview of the finished system with demo data; an installed site shows none. Search opens the command palette, like ⌘K.

What's new at the very bottom shows the version you are running, such as v1.0.6. Choose it to see what changed in each update, newest first. A release with a long title shows a short line while it is folded, ending in "and more"; open it to read the whole title. A dot on the icon means unread news.

The top bar

Roles

Reviewers, supervisors and administrators can also export clips and snapshots from recorded footage and move PTZ cameras. Viewers can watch recordings but cannot export them or move a camera. Licence, Settings, Apps, Users and Models are for administrators. A Display account is always a Viewer (chapter 17).

Custom roles. When the four built-in roles do not fit, an administrator can make a custom role: open Users, choose the Roles tab, choose Copy on a built-in role or on one of your own, give it a name (1 to 40 characters), an optional description and a home page, then switch on exactly what it may do. A role can never give more than you can do yourself, and a site can have at most 20 roles of its own. Built-in roles cannot be changed, only copied. Edit changes a custom role, and you cannot edit or delete your own role. To Delete a role that people hold, choose the new role they move to first. A custom role is picked for a person in the same lists as the built-in ones.

Themes, density and home page

The account menu at the bottom of the sidebar switches the theme and the row density (Comfortable or Compact), opens Your preferences and Notification settings, and signs you out. After sign-in you land on a home page that depends on your role: Overview for administrators and viewers, Monitor for supervisors and reviewers. Your preferences sets your own theme, row spacing, Date format, Clock (automatic, 24-hour or 12-hour) and Home page. A new date or clock choice applies to each page as you open it. They are kept with your account, so they follow you to any computer, and so does your Table or Cards choice on Events. An administrator also reaches it from the Your preferences panel in Settings → General & branding.

The screens move only when something changes. A page fades in the first time you open it and does not replay on a second visit or when you change a filter. Lists, charts, markers and counts move only when their data changes. If your system asks for reduced motion, a change shows as a short fade instead. A page opens already filled in after you rest the pointer on its menu entry. During a burst of events the lists on screen refresh together about every 1.5 s, while the alert count and pop-ups update at once.

Every page's help opens with a map of the page: a small sketch with numbered parts, what each part does, and what changes when you use its main controls. Thirty words the pages use, such as zone, grace, golden set and legal hold, show a one-sentence meaning when you point at them. Most pages, from Overview, Events and People to System and each camera's page, have Customize for the panels they show and what they open on, and every control says what it does, its default and its range. Press ? on any page, outside a text field, for a map of the keys that work there and everywhere; the account menu has it as well.

Cameras and zones

A camera is a video source. A zone is a part of that camera's picture where rules apply. Cameras are grouped into areas, such as a warehouse or a production hall.

Find cameras on the Cameras page searches the networks the server is on by itself, and lists the cameras that answer. A camera on another network or VLAN does not answer, so type its range in Look in this range, such as 192.168.1.0/24, and choose Look. A range runs from /24 to /32 (at most 256 addresses). In the Docker install the search finds nothing by itself, because the server sees only Docker's own network: type the range. Choose a camera from the list and type its login: the platform reads its streams (the password is never shown again) and opens the camera form filled in. A recorder shows Add all channels, which fills the Add from NVR form. A camera you already added is marked Already added.

Add or edit a camera

Open Cameras and choose Add camera. To change a camera later, choose Edit on its row, or open it and use its Settings tab. Administrators and supervisors can add and edit cameras. The platform stores camera passwords encrypted with APP_SECRET.

The Cameras page groups each camera under its area, and lists its source, status, checks and checks per second. Status reads Online, Starting, Offline, Disabled or No heartbeat: hover it for the last frame or error. It reads Password needs re-entering when the saved password can no longer be read (for example after the server's secret changed): the platform stops calling that camera until you type the password again. Changing a camera's port, or switching between rtsp and rtsps, also asks for the password again. If someone else saved the camera while you edited it, you get a "changed meanwhile" message. Deleting a camera removes its zones, rules and events, so you type delete first.

Test connection in the camera form asks the vision service to open the stream. It reports the codec, size and frames per second of each stream, or why it failed, and says whether the local or the outside address answered. When a sub stream is set, it compares the main and the sub stream. If they show different areas, the camera keeps its pictures and the form says why. The password is never shown. If the vision service is not running, it says "Vision worker offline".

A sub stream uses a second session on the camera, which the form notes beside the field. Use the sub stream on busy walls (Settings → Monitor defaults) keeps the walls off that second session.

Checks on each camera

Each camera row has a chip for every check whose app is on: PPE, Security, Safety and Counting. A solid chip is on and a dashed one is off. A check whose app is off, or not in the licence, has no chip. Switching a check on says what it adds to each picture, such as "Safety is on for Dock 2. Adds about 8 ms per picture (estimate)." Switching one off asks first, and says its rules are kept and apply again when it is switched back on. A viewer sees the chips locked, with the reason under them. Switching the Safety app off on the Apps page removes the Safety chip; the camera's stored Safety check comes back with the app. The Add or edit camera form has the same chips, under Checks.

Checking budget and slowed cameras

The Checking budget bar above the list shows how much of the server's checking time is in use, measured over the last 30 seconds. It turns amber at 80 % and reads Nearly full, and red over 100 % and reads Over budget. Its sentence says how much room is left, such as "Room for about 3 more cameras like these." The tag reads Measured once the server has measured its checks, and Estimate until then.

A camera's rate shows what it is set to. When the server cannot check it that often, the rate reads Checked 3.4 of 5 per second in amber, with the reason under it: the server is over its checking budget, or the camera sends fewer pictures than it was set to. Set to 3.4 beside it lowers the rate to what the server manages. The banner above the list covers every slowed camera at once with Set to what fits. Either change can be undone: a message offers Undo for a few seconds and puts the old rates back.

Running now

Running now (in the Cameras header; the same panel is on Apps and Models) lists each model the AI engine holds in memory: its milliseconds a picture, the cameras it serves, Change to pick another model, and its memory. A bar shows the server's memory, or the graphics card's memory where there is one, split into models, other programs and what is free. A model that does not fit in the memory kept free (scheduler.memoryReserveMb, Memory kept free under Settings → Detection, advanced; 512 MB by default) is not loaded. The panel says why in one sentence, such as "Pose needs about 0.3 GB, 0.2 GB free. Turn off Smoking.", and a check that needs that model shows the same sentence on the Rules page.

The licence limit

The licence sets how many cameras may run (Licence). At the limit, Add camera and Find cameras switch off and say "Camera limit reached". A camera beyond the limit is paused, not deleted: it shows Paused — licence limit in the list and on its page, and its Monitor tile says "The camera limit of the licence is used up, so this camera is not being watched or recorded." No live feed opens, and a paused camera raises no "is offline" alert. The newest cameras pause first, and the oldest keep running. An administrator can switch on Keep active on a paused camera in the Cameras list: it then runs, and the newest of the others gives up its place. Switch it off to return to the usual order.

Areas

Choose Add area on the Cameras page to create one. You can rename, reorder, fold and delete areas there. Deleting an area keeps its cameras and its rules: the cameras move to "No area". Two areas cannot have names that differ only in capital letters. While a search or filter is on, folding and moving areas is switched off.

Switch on Overlapping views for an area whose cameras can see the same people. The Monitor then counts people from the camera that sees the most and shows the most there could be, such as "3 up to 5"; it does not guess who is the same person. A violation both cameras see becomes one event. The event page shows the second camera's pictures under Also seen by. A person with no recognised id, or missing different items, still gets a separate event. The Area filter on the Monitor, the Overview and Reports narrows each page to one area.

An area also has Most people allowed, from 1 to 5000: a warning opens when more people than that are in the area for a minute (empty means no limit). Cameras see the whole area (off by default) says the cameras between them cover all of it, so the number inside is counted from the pictures, not from people crossing lines.

The Cameras page shows a banner for an area or camera that needs attention, and Customize changes how the groups show. Each camera's form has an area picker. On the Monitor, a camera you pinned to the wall stays pinned when you change the area, and Needs attention ignores the area.

Add cameras from an NVR

Choose Add from NVR. Enter the recorder's address, port and login, a vendor preset, the area and the first and last channel (1 to 64), then choose Find channels. The platform tests four channels at a time and unticks those that do not answer. Add creates every ticked channel, each with a whole-frame zone, in one step, or none of them: if a channel was already added, the dialog names it. This needs the vision service to be running.

Sources

Upload stored videos under Cameras → More camera actions → Stored videos. The platform accepts mp4, mkv, mov, avi and webm files of up to 4 GiB, and keeps a file a camera still uses. An upload shows a bar and a percentage. A stored video behaves like a live camera: it raises real events and records, so use it to test detection and rules.

Zones

Every camera gets a zone called Whole frame that covers the entire picture. Add more zones where the rules differ, for example a loading bay that needs a vest and a workshop that needs goggles as well. Zones can be drawn whenever an AI app is on, not only PPE Compliance.

1. Open the camera and choose the Zones tab.

2. Choose Add zone. Click on the snapshot to place the corners of the area. A zone needs at least 3 points and can have up to 64.

3. Press Enter, double-click, or choose Close polygon to finish. Press Esc to cancel. The platform creates the zone at once and names it "Zone 1", "Zone 2" and so on.

4. Or choose Click to outline and click once on a floor, walkway or door (or move the cross with the arrow keys and press Enter). A dashed outline is suggested. Accept creates the zone, Adjust lets you drag its corners first, and Reject or Esc drops it. Nothing is saved before you accept. Click to outline uses the SAM, SAM 2, MobileSAM or FastSAM model you activate on the Models page, from the next click and with no restart; MobileSAM is the light choice on a server without a GPU.

5. Select the zone in the list on the right. Type a new name in the Zone name field and click away to save it. Set its Grace there too.

6. To change the shape, drag the corners, or choose Redraw. Then choose Save, or Discard to drop the change. Refresh snapshot fetches a new picture from the camera.

Where zones overlap, the smallest zone that contains a person's feet decides which rules apply. Grace is how many seconds a person must be inside a zone before the platform judges them. The default is 2 seconds, and the range is 0 to 60. A zone you switch off with Enabled shows an Off badge. Deleting a zone deletes its rules too. If the picture area shows "NO SIGNAL", vision has no frame for that camera yet (chapter 18).

Fixed items

With the Safety Checks app on, a camera has a Fixed items tab. A fixed item is something that looks like fire, smoke or a forklift but never changes, such as a red lamp or a steam vent. The platform learns one when it keeps triggering a check in the same place; you can also choose Draw manually, type a Label such as "steam vent" and click the corners of the area to ignore. Each item shows a thumbnail and a status. Confirm keeps it, Reject says this place may really have a fire, so the platform never learns here, and Remove deletes it. Editing needs permission to edit cameras. The tab opens only while the app is on.

An item hides only a still copy of itself. While the platform is still learning it, and for the first 10 minutes, it raises only a low-severity event with no alarm. After that it ignores the item while it stays still. A check alarms again when the box grows, another box overlaps it, it flickers, or, for smoke, its top edge rises. An item not seen for 7 days expires. Reports show how many detections each camera's fixed items ignored (chapter 14).

Floor calibration

Every camera's Settings tab holds a Floor calibration panel, which turns positions and distances on that camera's floor into metres. The Person near forklift check needs it to measure in metres. Click four to eight spots on the picture that you can find on the floor, then give each its position in metres; Suggest points can place them for you.

Start from a rectangle

Measure a floor rectangle is one way to start: click the four corners of a floor rectangle you can measure, type its First side (corner 1 to 2) and Second side (corner 2 to 3), and choose Use these points.

Check the points

Each point shows its error, and a Tape check between two spots you measured gives a plain accuracy figure. Points that cannot be a calibration are refused in plain words.

Remove the calibration

Remove calibration says "Calibration removed: distances are approximate", and the forklift check goes back to the gap between boxes. Calibrations made before this update carry over.

Accuracy and size limits

A camera's own page (open it from Cameras) holds an Accuracy and size limits panel. Its first line gives the camera's accuracy grade and the day it was checked, such as "Accuracy: Counting": Counting (good enough for counts, presence and PPE), Presence and PPE only, Unreliable or "not yet checked". A grade taken from picture size alone says "(from picture size)". The grade comes from the accuracy checks (chapter 16). A second line says how much of the camera's zones is big enough to judge, such as "Usable area: 62 % of its zones. People are under 40 px in the rest."

The panel also holds size limits for that camera alone: the smallest person, in pixels tall, each check accepts. When the camera's own check set suggests a presence limit, "Suggested from this camera's check set" offers Use it. Save size limits needs permission to change cameras.

Each day the platform compares a camera's picture with a saved one. If the view has shifted, or the camera's stream changed, its accuracy reads "not yet checked" and an alert asks you to check its zones and collect a fresh check set. Before a new model, a counting setting or a camera's size limits takes effect, it is replayed on the camera's saved check sets. A camera whose new figures do not hold up loses its grade at once, and saving without that check also leaves the camera ungraded until it is checked again.

Counting lines

With the People Counting app on, a camera has a Counting lines tab. Give a line a Name, such as Main gate, and click 2 to 8 points along where people walk through; the arrow points the way people are counted in. Counts into area and Counts out of area name the areas people enter and leave when they cross in. Each line shows today's counts, such as "Today: 42 in, 39 out", and can be switched off. The error of these counts is not measured yet. The line settings, such as Smallest person counted at a line, are under Settings → Counting & coverage.

Counting page

With the People Counting app on, Counting (Operate; the page title is People counting) shows who came in and went out, and how many people are inside each area now. Where the cameras of an area overlap and are joined on the floor map, a person two cameras see counts once and the figure is one number; if they are not joined, it stays a range and the row says why. Four tiles lead: Counted in, Counted out, Inside now and Busiest hour. Below them are a bar for each hour of people coming in and going out (a dash is an hour no counting camera was watching), Inside now for each area against its Most people allowed limit, and a table of Counting lines. An area over its limit for a minute raises an alert.

• Period and Area. Choose Today, Yesterday, Last 7 days or Last 30 days, and one area or All areas.
• Download. CSV, per line, CSV, per hour or PDF report, for the period and area on screen.
• Limits of the figures. A line checked against a careful count shows how far off it can be, and its tile reads like "1,240 ± 74" with how many lines were checked. A line not yet checked keeps its count and says "Error not measured", and so does any total it enters. Crossings the platform could not place show as a higher end ("up to 1,262 with 22 unsure crossings"), a line that tends to read low or high says so, and a line off by more than 20 % stays in the table but leaves the totals and is named under them. Count crowds is Planned. Until a line exists the page says "No counting lines yet" and offers Open Cameras.

Vehicles

With the Vehicles & Plates app on, Vehicles (Operate; the page title is Vehicles and plates) counts cars, motorbikes, buses, trucks and bicycles crossing a line on a camera, in and out. It works with or without PPE detection, but it needs a person model that can see vehicles: if the one in use cannot, the page says "Vehicles cannot be counted yet" and links to Models.

• What it shows. Vehicle crossings, Wrong-way alerts and Vehicles by type, for Today, Last 7 days or Last 30 days, for every line or the one chosen in Line. Each pass over a line counts once, so in plus out is crossings, not vehicles.
• Draw a line. Administrators and supervisors choose Set up lines and bays, pick a camera and draw a line across a gate or road on its picture. Allowed direction is Both ways, Only in or Only out, and a vehicle going the other way raises a wrong-way alert. Switch Ring on wrong way off and such a vehicle is still logged and counted, but no alarm rings. Parking bays are drawn in the same place: choose the Parking bays tab, and the page shows each bay taken, free or reserved.
• Boxes and reports. Playback draws each vehicle's box and type on video from a camera with a vehicle line. Reports → Vehicles gives crossings by type and hour on each line, and every wrong-way alert, as a landscape PDF or a CSV.
• When it cannot count. A dash is not zero: it means nothing was checked, and the page says why, for example "Camera offline since 10:12". "No vehicles in this period." appears only for a period the platform watched and saw none.
• Events. A wrong-way crossing appears in Events as Wrong way (Medium severity), with an alarm unless it is switched off for that line.
• Settings. Settings → Vehicles sets Smallest vehicle counted (60 px wide). A bay is taken after (20 s) and A bay is free after (60 s) set how fast a bay changes. Settings → Vehicles also holds the plate layouts: add the layouts your site uses, such as AA AAA-9999 (A is a letter, 9 a digit, # either), and each read is checked against them; with none listed, any plate text is accepted.
• Number plates. The platform can read the plate of each vehicle that crosses a gate line, on the computer at your site. First install the Plate reader card on the Models page and activate it (Latin letters and digits only); until then Read plates here cannot be switched on. Then, on a vehicle line in Set up lines and bays, switch on Read plates here (off by default; it uses the camera's main stream, so it needs more detail and more computer time). A plate is accepted only when several pictures of the vehicle agree, and a faint or too small plate is shown as Plate not read.
• Gate events. Each pass with a plate appears under Gate events. A plate on a deny list, an unknown plate or a near match shows Plate waiting for a guard until someone chooses Let in or Turn away. That choice is recorded and moves no barrier. A reviewer can correct a misread plate. While a plate waits, the Monitor's automatic layout brings that gate camera forward.
• Plate lists. Plate lists holds an Allow list, a Deny list and a visitor list, with groups, end dates and reasons, and can be imported or exported as a spreadsheet (CSV).
• Who sees plates, and privacy. The permission See number plates (administrators, supervisors and reviewers, not viewers) shows plate text and pictures and allows the plate search; without it a plate shows as Hidden. Edit plate lists changes the lists. Plate pictures are kept for 30 days by default, and a plate on no list is kept as text unless you turn that off. A plate search, an opened plate picture, a list change, an import and an export are written to the audit log as a hash of the plate, never the plate itself.
• Erase a plate. An administrator can erase one plate everywhere: its passes with their pictures, its list entries and the plate on events, where the text becomes "Erased". An event under a legal hold keeps its plate, and the answer says how many were held back. A privacy request about a plate runs the same erase and exports that plate's data. The event's own snapshot may still show the plate.

Conveyor

With the Conveyor Counting app on, Conveyor (Operate) counts the items on a belt with the camera above it. Each belt line shows the Batch count, the Total and the batch number. The batch count goes back to 0 when a new batch starts. The total keeps going.

• Draw a belt line. Administrators and supervisors open the camera, choose its Counting lines tab and, under Belt counting, draw a line across the belt, then a belt area that leaves out the belt edges. The line needs an Object model that can find the item; the page says so when it has none.
• New batch. Press New batch when a new lot starts. A second press within 10 seconds is refused, so a double tap starts one batch. The button needs permission to review events. The operator panel's own New batch button does the same.
• Open count screen. The button opens the count in large numbers for a monitor. A display account sees it with no New batch button. Not live means the number may be out of date. Not counting means the camera or the analysis is stopped. Do not trust the number until the label goes away.
• Batch report. The Batch report menu gives the batches of a period as CSV data or a PDF report. Recent batches lists the latest ones. Enter the items counted by hand for a finished batch with Enter (or Change), and the page shows the measured error. Until batches have been hand counted it says "Error not measured".
• Operator panel. Settings → Operator panel (administrators only) sets the Panel link (Off or Serial), the Serial port, Speed, Device number and the Register map: the registers that hold the batch count, the total, the batch number, the status and the New batch request. Check the map on the panel before you switch the link on. The panel keeps its last number while the computer or the analysis is off; it is not live then. A cable problem never stops counting, and the page says the link has a problem.
• What is kept. Every counted crossing is kept for 7 days, so a resent batch of counts is never added twice. Batches are kept. Neither holds pictures.

Floor map

Floor map (Operate) shows the site as one place: where the cameras are and what they see. It opens in 3D, with 2D and a flat plan when you need them, and a Floor list when the site has more than one floor. With no map it says "No floor map yet" and offers Start a map. Administrators and supervisors also see Edit in the studio, which opens five steps. Until a camera is calibrated, people show as a count on that camera.

1. Plan. Choose Choose a file for a picture or a PDF page of the floor plan, or Start drawing. The tools are Select, Wall, Room, Stairs, Lift, Fixed item, Door and Label. Mark a line you know the length of to set the scale. Measure again at any time and the whole draft is rescaled. A draft with walls, doors or cameras already drawn on an unscaled picture is rescaled with the picture when you set its first scale.

2. Cameras. Pick a camera and click the plan to drop it, choose Aim: click where it looks and click, and set how far it sees. A camera already on another floor says so and cannot be dropped twice. Take off the map removes one.

3. Calibrate. The camera's picture sits beside the plan, so clicks on the plan give floor points in metres. A camera calibrated on its own can be fitted to the plan with Fit to the plan, by dragging its outline or clicking two matching points. Calibrate again on the plan starts over.

4. Zones. Choose Open the Zones screen to draw a zone once and have it apply to every calibrated camera that sees it.

5. Publish. Choose Review and publish, then Publish to the live map. Nothing on the live map or in detection changes until you publish.

A camera on no floor map shows Not on the map in the camera list, with a link to place it. Editing the map needs a screen wider than a phone's. A photo of a hand-drawn sketch can be traced into a draft map in the map editor.

Trails and path replay

The cameras follow each person on the floor map and keep a trail of their points: about one a second while they walk, and one every 10 seconds while they stand still. Choose Replay this path after you pick a person on the live map, or Replay in the header of the Floor map page for the whole floor. The replay page draws the route across the map, with the matching clips beside it and a timeline below.

• Gaps. The line is solid where a camera saw the person and dashed where none did. A gap of more than 5 seconds is dashed and the timeline says how long the person was not on camera. A person who stood still, with points at most 15 seconds apart, is not shown as a gap. With Skip gaps on, playback jumps over the dashed time.
• Handoffs. Each move from one camera to the next is a handoff, marked with an open circle. A handoff with a link score under 0.7 is only possible. A reviewer chooses Confirm or Reject; a rejected handoff is dropped, and the route is no longer drawn solid across it.
• Playback. Play or Pause (Space), the left and right arrows step between sightings, and the speed is 1×, 4× or 16×.

Trails need permission to see people and the People tracking module in the licence. Each trail point is kept for 30 days at most, and is deleted with the person (see the security sheet).

Who was in

Below the map on the replay page, Who was in asks who was in a room between two times. Choose the Room, then the From and To times. The page lists each confirmed person with a Replay path link, and counts the possible ones: "more possible tracks: check it in replay". It also says that people in blind spots or off camera are not covered.

Evidence packs

Evidence pack on the replay page makes one zip of a path. Untick what you do not need from Picture of the trail, Timeline file, A clip from each camera and A snapshot from each camera; each shows its file count and size. Type a Case reference (up to 60 characters) and a Reason, then choose Make the pack. The export is recorded with the case reference and the reason.

When the pack is ready, the dialog says "Manifest written" with the number of files, lists each file with its SHA-256 check value, and names any clip no recording covered under "Not in the pack". Choose Download the pack to save the zip. Making a pack needs permission to see people, permission to export recordings and the People tracking module.

Heat layer

The Heat layer colours each square of the live map for the past 24 hours. Choose one measure: Violations, Traffic (how many different people passed) or Time spent (how long people stayed). Traffic and time spent need permission to see people. The map draws at most 2,000 squares, the busiest first.

Muster and roll-call

Choose Muster on the Floor map page, then Start a muster. The dialog starts a drill or a real muster, with its muster points. Muster points shows how many people reached each point and how sure that count is. A floor with no muster point drawn says so and counts nobody.

Each person not yet accounted for is listed under the room a camera last saw them in, in Last seen inside, room by room, marked Not yet counted. Mark safe and Cannot find record what you find; they need permission to review events and to see people. At the close, anyone still open reads Unaccounted at close. Print prints the roll-call.

Picture health

Cameras, then Picture health, shows each camera's picture against the reference picture a person saved for it. Each camera has one state: Good, Blurred, Blocked, Moved, Dark or Frozen. The worst cameras come first. Checks run once a second, and a state is reported once it has held for 10 seconds.

• Save a reference. Choose Set a new reference and confirm. Only a person can save one, and it never refreshes itself. A blurred, blocked, dark or frozen picture cannot be saved.
• Thresholds. The Day threshold (30 % by default) and the Night threshold (45 %) say how far the picture may differ from its reference before an alert. Each one is set between 10 and 80 %, in steps of 5.
• Infrared. The night threshold applies while the camera is in infrared.
• Alerts. A picture problem shows on Overview, on the Monitor and in the bell, as "camera name picture: state".

Rules and PPE types

A rule says that one PPE type is required in one zone. Each rule has a severity and an optional schedule. A zone with no rules checks nothing, and the Rules page marks it "Not checked".

Set requirements

1. Open Rules and stay on the Requirements tab. It lists every zone of every camera, with one column for each enabled PPE type.

2. Click a cell to require that item in that zone. Click again to remove the rule.

3. Use the arrow in a cell, or right-click it, to open its settings beside the table: Required, Severity, Active hours, Days, Warn after (seconds), Violate after (seconds) and Paused, which keeps a rule but stops checking it.

Many zones at once

• A cell's settings offer Require everywhere, Set all to its severity and Remove everywhere for that PPE type.
• Tick rows, then use Copy from…, Apply preset… or Clear rules. The presets are Basic site (helmet, vest, boots), Battery / chemical area (helmet, goggles, face mask, gloves, boots, coverall) and All enabled PPE types. A preset applies only the types that are enabled.
• The menu on a zone offers Copy this zone's rules to…, Import rules from…, Edit details…, Open camera and Clear all rules.
• Copying, importing and presets replace a zone's rules. They do not merge with them.
• Arrow keys move between cells, Space toggles a cell, Shift F10 shows its settings and Esc returns to the cell.

Each camera's Rules tab shows the same rules one zone at a time, with a Required switch per type. Reviewers and viewers see the page but cannot change it.

Severity

Every rule is High, Medium or Low. An event takes the highest severity among its missing items. By default only High pops up (chapter 12), and Medium and High count in the alert number.

Schedules

A rule can apply only at certain times. In the cell's settings, set Active hours and the Days of the week. Set both a start and an end time, or leave both empty to check all day. Hours may cross midnight, such as 22:00 to 06:00. Leaving all seven days on means the rule never pauses.

Warn after and violate after

Each rule can set two times. Warn after (0 seconds by default) is how long a person may miss the item before they show as a warning. Violate after (the site's violation time, 8 seconds by default) is how long before an event is raised. Leave a field empty to use the default. Warn after cannot be later than violate after. The Rules page groups zones by the area their camera is in.

PPE types

Open Rules → PPE types to see the gear the platform can check for. A disabled type is ignored in every zone. The first six types are on by default. The last six are available but off, until you switch them on and have an active model that detects them. Start from a preset, beside Add type, applies a ready-made set of types. It names what it will switch on and off before you confirm. Construction and industry is Safety helmet, Safety vest, Safety boots, Gloves and Goggles. Hygiene & food safety is Hairnet, Face mask, Gloves and Apron, for kitchens and food plants. A preset re-creates any of its types you deleted, keeps every zone rule, and is recorded once in the audit log. If the PPE Compliance app is off, it answers "The PPE app is switched off." No ready-made model detects hairnets or aprons yet, so train one or map a model's classes to them.

Choose Add type to create a custom type. Give it a key (lowercase letters, digits and underscores, starting with a letter), a label, a colour and a body region. A new type starts enabled and is off in every zone until you require it. Pick a key that matches a class your model predicts, or map that class to the type in the model's class map (chapter 16). Each type's menu also offers Edit, Require in every zone and Delete. Deleting a type stops every rule that requires it.

Safety checks

Rules has four tabs: Requirements, Safety checks, Security checks and PPE types. Open Rules → Safety checks to switch other checks on for each zone. Zones run down the side and the checks across. A check shows only when its app is switched on and in the licence (Apps); when none is, the tab says No safety checks are available. Switch a cell on, then use its sliders button for Severity, Alarm (siren and alarm channels), the check's own values and Active (empty = always) hours and days. A refused change reads as a plain sentence. A check that needs a model nothing has loaded shows Needs a model · Get model, a link to Models. Each check has its own colour, used on the Monitor, in Events and in Reports.

A check raises an event only after it holds through its confirmation window, never on a single frame. Where a check lists it, Confirm for and Seen in at least set that window. A pose model needs the PyTorch engine, except RTMPose.

Events from these checks appear in Events with a Check filter (chapter 11). Fire and person down raise alarms (chapter 12). A lamp, a vent or a sign that looks like fire keeps triggering a check: the camera's Fixed items tab teaches the platform to ignore it (chapter 6).

Security checks

Open Rules → Security checks for the checks of the Security app. It works like the Safety checks tab, and says No security checks are available when the app is off or not in the licence. A check that cannot run yet says why under its name, for example "Needs the People & Visitors app. Switch it on in Apps."

The three face checks need the People & Visitors app and Recognise named people by face (chapter 8). Most security alerts per camera per hour (Settings → Detection, 60 by default) stops a camera flooding the bell: over the limit, the camera keeps its further alerts that hour as sightings and shows one notice. Alarms still ring, and 0 means no limit.

A face alert shows the face that was seen beside the enrolled photo, how wide the face was against the rule's smallest face, the reason a person is on the watchlist, and the same person's earlier alerts today. Alerts raised before the width was recorded show no width.

PPE and security modes

The platform can check PPE, or it can only track people. You choose per camera, and you can switch PPE checking off for the whole site.

PPE on each camera

Whether a camera checks PPE is its PPE chip, set on the Cameras page or in the camera's form (Checks on each camera).

• PPE on is the default. The camera checks the PPE each zone requires and raises PPE events.
• PPE off makes the camera track people only. It keeps its zones and PPE rules, which apply again when PPE is switched back on, and no PPE events are raised. Its other checks, Security, Safety and Counting, carry on.

The site-wide PPE checking switch

Administrators can switch PPE checking off under Settings → Detection → PPE checking. It is on by default. The platform asks you to confirm before it turns the switch off. A camera judges PPE only when the site switch is on and the camera's PPE chip is on.

What happens to a camera that does not judge PPE

• Tracking, the live view and recording carry on. There is no PPE check, no compliance verdict, no frame mining and no PPE event.
• An event that was open when the PPE chip or the switch changed is closed. It stays on record.
• When no enabled camera judges PPE, the platform unloads the PPE model and frees its GPU memory. It loads the model again when a camera needs it.
• A change reaches the vision worker within 15 seconds.
• Nothing is deleted. Zones, rules and past events stay, and switching PPE checking back on restores every camera's own PPE chip.
• The Test bench still works while the site switch is on, even if no camera has its PPE chip on. With the switch off it answers "PPE checking is switched off". Model downloads and training are not affected.

Face recognition

Settings → Detection also holds Recognise named people by face. When it is on, the platform recognises the people you have named, on any camera and on later days, whenever their face is visible. Masks and respirators hide it. It works only when Follow people across cameras is on (Settings → Re-identification). The first start downloads two small face models, about 38 MB. A site without internet then matches people by appearance only.

Keep unknown faces for sets how many days the platform keeps the face templates and thumbnails of people who have no name, and 0 keeps none. Under the strictest privacy rules it is 0, so the face of a person with no name is not kept; a profile that allows more sets the limit. The face data of a named person is kept until no camera has seen them for a year: Settings → Privacy → Delete face data of people unseen for (Advanced) sets the days, 365 by default, and then the face photos, templates and consents are deleted and the name stays. Two weeks before, the System page warns how many people will be deleted and on which day, and a person on legal hold is skipped.

Settings → Privacy (administrators) holds the site's privacy profile. Until an administrator picks the site's country, the strictest rules apply. Picking a country applies its profile at once; the dialog first lists every value that goes down, and values never go up by themselves. Each setting with a limit shows it beside the value, and a value above the advised one asks for a reason, which goes into the site's privacy paperwork and never locks anything. Cameras record no sound unless an administrator switches on Record sound for one camera, and a profile can keep that switch locked off. Fill the privacy contact and the sign-in page names it.

Watching live

Open Monitor to watch your cameras as a wall of tiles. Supervisors and reviewers land here after they sign in. A line under the title gives the cameras on screen, the people in view and the compliance right now. The summary's compliance card shows Time in compliance, last hour with its coverage and best and worst case, compared with the hour before (how the numbers are counted).

The toolbar

Your camera choice and layout are saved in your browser and in the page address, so you can bookmark a wall or send its link. For a second display, open it with ?display=wall.

Live video and Boxes

Each tile plays the camera's own video at its own speed, with the platform's boxes, labels and zones drawn over it.

Boxes

Boxes follow every picture the video shows and glide between checks (5 a second by default). When the detector misses a person or vehicle for a moment, the box stays on its predicted spot for up to a second; a held box is only drawn, never counted or judged. Boxes shows or hides them at once, without reloading; Boxes on all above the wall, or B, does every tile.

Streams

Small tiles and big walls play each camera's sub stream, and Focus and large tiles play its main stream, so a wall of many cameras does not open the main stream for every tile. Tiles change without a black flash. Stored videos and webcams show pictures.

Outside the local network

A viewer whose browser cannot reach the camera's video directly, such as someone outside the local network, gets the video through the server instead, about a second behind, and only then falls back to pictures. On an iPhone this needs iOS 17.1 or later. A tile whose video cannot play says why and retries.

Reading a tile

A badge on every tile says what you see, such as "Live video · boxes 5/s"; point at it for the stream and picture size, or for "Video through the server". The line under the wall sums it up. Point at a person's box and their row lights in the summary and the track list.

Boxes on recorded video

Playback draws the same layer over a recording. Above the tiles, Boxes shows or hides it, Labels names each box and says its verdict, and one button per PPE type shows or hides that type. A chip says how often the boxes update, such as "Boxes update 5 per second". The recording plays at its own picture speed, and the playhead moves at the speed you choose, from 0.25× to 8×. A camera whose video was not checked has no boxes, and the chip says so. The system saves what it found beside each recording, about 2 MB an hour for each camera, and removes it with the recording, whether the automatic clean-up or a clean-up you start by hand in Storage deletes the video. A camera with a vehicle line also shows each vehicle's box and type. Names and person ids stay hidden from anyone without access to people. An export shows only what the system found at that moment, never the smoothed boxes shown while playing.

Reading a tile

A thin strip on the tile's left edge shows the camera's state: Online, Starting, Offline or Disabled. The look of the labels depends on the view you choose. This table describes the default look, called Reference. The colour of an item box is the colour of its PPE type, listed in chapter 7.

A person who is missing an item but has not yet reached the violation time is a Warning: the tile's tab, the track list and Active alerts say so in amber, and the counts card has a Warning row. The words always name the item, so a warning is never told from a violation by colour alone.

By default the Summary panel shows severity as a pill: ALERT for High, WARN for Medium, NOTICE for Low and OK for a compliant person. A chip on each tile counts its people: 1 missing, or 2 OK. Where cameras overlap, the people count reads as a range, such as "3 up to 5": the camera that sees the most, then every view added together. Where the overlapping cameras are all placed on the floor map and can be joined, a person seen by two of them counts once, and the range closes to that one number.

Customize view

On the Monitor, Customize opens beside the wall on the left (from the bottom on a phone), so each change shows at once; on Settings and the Test bench it opens as a window. A few parts always stay on, such as Needs attention and the line saying how a figure is counted, and each says why. It has seven tabs: Presets, Boxes, Labels, Colours, Summary column, Auto layout and PPE wording. Every option has a plain name and one line saying what it changes. Your look follows your account to any computer. The presets are the quickest way in:

• Reference: full detections with a compliance tag on each person.
• Control room: amber violations on corner boxes, with no icons. Calmer for a wall display.
• High contrast: thick boxes and large labels for a screen viewed from across the room.
• Minimal: thin corner boxes, no OK labels, no clock and no Summary. Only the violations.
• Detections: every detection in its own colour with its confidence.

The Boxes tab also sets the overlay style: Detections, Compliance or Both. Skeleton (pose models) is off by default; switched on, it draws the body outline on each person when the camera runs a pose model. A change shows at once and is kept with your account, so it survives a reload and another browser. Reset to default puts back the site default, with Undo for 7 seconds. Save as site default is for administrators only.

The Auto layout tab has a separate switch for each thing Auto may do by itself: move tiles, bring in a camera from outside the wall, and fill a wall screen for a high event, for a number of seconds you choose. It also sets which events move a tile and how long a camera stays in front; those rows need Move tiles switched on. Beside the wall, the panel also sizes Needs attention Standard or Narrow (it always stays on, and says why), sets Row spacing, and has three switches for the lines around the wall: the layout caption, Why did it move? and the tips.

Playback and export

Open Playback (Operate) to look back at what the cameras recorded. It shows up to four cameras at the same moment: one main picture and up to three small ones. Choose a tile to make it the main picture. Double-click the main picture, or choose Fill, to fill the grid with it; Show all four brings the others back. A tile says so where its camera recorded nothing at that moment. A recording appears about a minute late, because the minute being recorded is not ready until it closes. Anyone who can read cameras can watch, viewers included. On a phone Playback shows one camera: swipe sideways to change camera.

The play bar has Play (then Pause), a Speed list from 0.25× to 8×, where 8× plays 8× key frames only, buttons that step to the previous and next event, and the playhead's clock on the site's time. Space plays and pauses, the left and right arrows move 1 second, Shift with an arrow moves 10 seconds, and B switches the boxes. Sync cameras keeps every picture on the same moment; switched off, only the main picture moves. Pause at events stops playing at the next event mark. Customize sets whether the page opens with it on. Boxes, Labels and a button for each PPE type draw the system's boxes over the recording (Boxes on recorded video).

The timeline shows the last half hour, with one lane for each camera. A solid bar is time that was recorded and a hatched bar is time that was not. Marks show motion and events: point at an event for its picture, or click the timeline to move the playhead. The Day strip counts today's events for each hour and boxes the 30 minutes in view; choose another day to move there. Type a time such as 10:38 in Jump to time to move the playhead. The Legend explains the marks.

Customize on Playback shows or hides Layer controls, the Day strip, Jump to time and the Legend. It also sets what the page opens with: Boxes and Labels (On or Off), Speed (1× unless you choose another), Sync cameras and Pause at events. The recorded and not recorded bars, the chip that says how often boxes update and the manifest line in the export panel always stay, because they say what exists as evidence and how exact it is.

An Open in Playback link on an event page and in an incident opens Playback at that moment, and a Look back 5 minutes link in the Monitor's camera panel opens it five minutes before now. A closed visit has Play back this visit. Only roles that may watch recordings see these links. When a camera has saved views, Playback lists them under Saved views, each as a link to the live picture.

To export, choose Clip, MP4 or Snapshot in Export type, set From and To (or drag the handles on the timeline), tick the cameras and choose Export clip or Export snapshot. A camera with nothing recorded in the range is unticked and says Nothing recorded in this range. With Boxes on, the export holds the boxes the system found at that moment, not the smoothed boxes shown while playing. If people may be in view, a Reason for this export of at least 4 characters is asked and logged. No reason is asked when nobody is in the pictures. On a phone the export panel opens as a sheet from the bottom.

The zip holds one video (or picture) per camera, and a manifest.json with the time range, who exported, the reason and a SHA-256 hash of every file. With Boxes on it also holds each camera's boxes. Verify hashes the files again and says how many still match. Choose Download within 14 days: after that the files are deleted and the export says "This export's files were deleted." Exports use the Exports category on the Storage page. Only you or an administrator can see your exports. Exporting needs the export permission, which viewers do not have.

To add a microphone on the Sound page, pick a sound card or a camera whose video carries sound. A camera's own microphone needs no address or password and also sets the camera that sees where it hears. Its sound is read in memory for alarms only; recordings and live views still carry no sound unless Record sound is on.

Cameras keep pictures only unless an administrator switches on Record sound for one camera (chapter 6). That camera's sound is then kept in its recordings, evidence clips and exports, and live view and Playback play it muted until someone unmutes it with the button on its tile. Every other camera's recordings, clips, exports and live view carry pictures only, even from a camera with a microphone. Recordings made before sound was dropped keep their sound until they age out, at most 30 days later.

On the Monitor, in Focus, a camera that turns shows a Live camera control pad: hold a direction button to move (a tap moves it a moment), let go to stop, use the zoom buttons, and choose Home position or Go to this view under Saved views. To save a view, type a name of up to 32 characters; the view's Remove button offers Undo. Moving a camera needs the move-cameras permission, which viewers do not have; their pad is disabled and says "Your role cannot move a camera. Ask an administrator." Saving a view needs permission to edit cameras.

Share safely

To share a clip outside the site, choose Cover faces and screens on Playback. Share safely finds the faces, screens and people for one camera and a range; you watch the whole clip, pick what to show, say why you are sharing it, and export. Every face, screen and bystander you do not show is covered with a solid fill, and the export holds only that covered, silent copy, signed. The original recording stays on site and nothing is recognised or matched.

Personal text. The text that travels with a clip can name people too. Names the site already knows (users and named people), emails, international phone numbers and the local ID card and phone formats of the site's country are found in the reason and notes you write and replaced by a mark: [NAME], [PHONE], [EMAIL] or [ID]. The export panel has a Personal text group that lists each find with Keep or Mask, so you can keep one as written when it is meant to be seen; a role that cannot see people cannot keep any. The same applies to the notes and comments in an incident's record. Downloads from Reports (PDF and CSV) have a Hide personal details switch for their free text, and after a download that replaced something a line says how many personal details were replaced by a mark. The audit log keeps the reason as written, because it stays on site. Names are matched as whole words of at least three letters, so a name can be missed in a different spelling; read the text before you send it.

Every export is signed with this site's own key unless signing is off, so whoever receives it can check that nothing was changed. Settings → Sharing holds Sign exports (on by default), this site's key with its fingerprint to give recipients, and Make a new site key (administrators). In the export panel, Sign this export can leave one export unsigned: a ready export says "Signed with this site's key" and the fingerprint, or "Not signed. The file carries none of this." A signed zip also holds a list of check values, the signature, the public key and a HOW-TO-CHECK.txt with the steps and commands to check it offline with OpenSSL, so a recipient needs no account. An export made before signing was added stays as it was.

Open Share safely (Operate; it needs the right to export recordings). Under Check a file you can test a file someone sends back. Choose Choose files, or drop one file, several files or a whole export zip: nothing is uploaded, only the check values are sent. The verdict reads Unchanged, Changed (naming the file that is not the one exported), Signature does not match, Signed by another site or Not known here, and ends: "This shows the file is unchanged, not that the scene was real." Recent exports lists this site's exports: their files are deleted after 14 days, but their record stays, so a copy can still be checked.

How a violation is decided

Nothing fires instantly, on purpose. For each tracked person and each required item, the platform keeps a smoothed probability that the item is present. It opens a violation only when the evidence stays bad for a few seconds.

1. The person has been tracked for at least 2 seconds, has been inside the zone for longer than its grace period, and is taller than 80 pixels.

2. The item's probability stays below 0.35 for the violation time (8 seconds by default) without a break. The violation opens, and so does an event. When the model can tell a bare head from a hidden one, only a bare head counts: a helmet out of sight pauses the clock, and the live view says Can't see and the item.

3. The violation closes only after the probability stays above 0.6 for 2 seconds. The gap between 0.35 and 0.6 stops a borderline item from opening and closing events again and again.

4. The same person cannot open another event for 60 seconds after one closes.

Two more values set how sure the detector must be before it counts a box: personConf for people (0.4) and ppeConf for PPE items (0.25). Administrators edit all of these under Settings → Detection → PPE evidence. The zone grace is set in the zone list (chapter 6).

Between a rule's warn after and violate after times, the person shows as a Warning in amber, with the item named in words. Nothing is saved until the violation time is reached.

One event covers one episode: the stretch of time in which one person is missing PPE. If more items go missing during the episode, they are added to the same event, and an event that is already open adds no second violation. A snapshot taken in the warning stage says Warning. An event records snapshots when it opens, and a clip of the episode with 5 seconds before and after (clipPreS and clipPostS under Settings → Retention). The clip is ready about 65 seconds after the event ends, once the 60-second recording segment that holds the end has closed. If recording was off, the event has snapshots and no clip.

A person hidden for a moment keeps their open event, so a pillar or a passing truck does not create a second one. When a camera loses a person briefly and finds them again, it joins the two pieces of their path into one person with one violation. Pieces it is not sure about stay apart and count as possible repeats; the joining scores, such as Score to join two pieces, are under Settings → Counting & coverage. When two people pass close enough to be mixed up, both are checked again for 2 seconds. An event ends when its violation was last checked, not when the person leaves view.

Events and review

Every confirmed violation becomes an event. Open Events to see them all.

The list

Switch between Table and Cards. The columns are Frame, What, Camera, When, Severity, Alarm, Review and Person. What names each event by its check and subject, in the check's colour; a PPE event still lists the missing items.

Filters

The filters are Camera, Check (starting at Any check), Missing PPE, Severity, Review, Person (a name or an id), From, To, Open only, Not acknowledged and Held only (events under a legal hold).

Export

Export CSV saves the list as it is filtered, the Check filter included.

Times

Every time on Events, and the From and To filters, follow the site's clock. An open event of a camera silent for 3 minutes is closed at its last sign of life and left out of duration figures.

Who sees the Person column

The Person column and filter are for admins, supervisors and reviewers; a viewer's list has neither.

Quick filters

Review set to unreviewed lists what still needs a decision. Open only lists events still in progress. Clear filters resets them.

Bookmark and page size

The filters and the chosen event are part of the page address, so you can bookmark or share them. The list shows 25, 50 or 100 events per page.

Open and closed

An event is open while the person is still missing PPE, and closed once they are compliant again or have left the picture.

One event

Choose an event to open it, beside the list on a wide window or on its own page on a narrow one. Above the picture, two lines say what the event is and where: for example, "Something new in Whole frame differs from its reference picture" and "Outlined: where the picture differs from the reference." Toggle Annotated and Raw to see the snapshot with or without the platform's boxes. Pictures of old events show only the items the camera's rules asked for. A clip plays in the same frame, with its marks moving with the video, or the page says why there is none: not cut yet, recording off, extraction failed, or unplayable. A clip made before this change has no saved marks, and the page says so. Open in Playback opens the recording at the event. If the disk was full and the snapshot could not be saved, the picture says Snapshot not saved (disk full).

Open full screen (on the snapshot or on the clip bar) opens the snapshot or the clip over the whole screen, with the same marks, and zoom, pan and pinch on a phone. On a desktop, Esc closes it, ← and → switch between snapshot and clip, Space plays the clip, B switches Annotated and Raw, and +, − and 0 zoom. On a phone, swipe left or right to switch between snapshot and clip. Closing it returns you to the button you opened it from.

When the camera kept no recording of the event's start, the clip is not there. The page then says: "No recording of that time is kept." It does not claim that recording was off. A supervisor or administrator also sees a link to the camera's recording setting.

When an area has overlapping views and two of its cameras saw the same person miss the same items, the event page has an Also seen by section with the other camera's picture. Delete event removes those pictures too.

The Details panel lists the camera, zone, start time, duration, severity, missing items, person, review state, absence score, track number and model. Absence score is how clearly the missing items were absent. It is a score, not the chance that the event is right. PPE probability at trigger shows each item's probability when the event opened.

Alarms

An event from a check that raises an alarm, such as fire or person down, has an Alarm column in the list and an Alarm card on its page. Choose Acknowledge (on the page, Acknowledge alarm) to show that someone is dealing with it. It needs permission to review events. An alarm is acknowledged once; afterwards the list and the page show Acknowledged, with when and the note. Acknowledging switches the siren off (chapter 12). Evidence cards, the bell and its toasts name the check too.

Reviewing an event

Reviewing needs the Admin, Supervisor or Reviewer role. Read the snapshot, then choose one of two buttons or press its key. The event shown stays put while new ones arrive at the top. You can add a note of up to 2000 characters first.

• Real violation (T): the person really lacked the item.
• Not a violation (F): the platform was wrong. The person wore the item, or the box was not a person.

Marking an event Not a violation also puts the raw frame, with the model's boxes, into the Datasets review queue, so the next model learns from the mistake (chapter 16). If an event is wrong because of the zone or the rule, fix the zone or the rule instead of marking it false. Your reviews feed the Reviewed so far panel on Reports. It counts the events a person has looked at, not accuracy.

Deleting events

Only an administrator can delete events. Delete event is on the side panel of Events (on the event page on a narrow window). It asks you to type delete, then removes the snapshots, the clip and its marks for good. It does not run for an event on legal hold, for a person on legal hold, for an event still running, or for an event whose pictures a training sample uses; the page says why.

Clear events, in the Events header, deletes many events at once. It works on the ticked rows, or on everything the filters match. Events on legal hold, still running, confirmed as violations, open incidents or used by a training sample are skipped and counted, and the message gives each count. Events that arrive after the dialog opened are not deleted. Every request is written to the audit log.

Legal hold

Put an event on legal hold from its page to keep its evidence. A held event is never deleted automatically, whatever the storage age rules say, and neither Delete event nor Clear events removes it. The Held only filter on Events lists the held ones.

Incidents

Incidents works violations as cases. Each incident has one status: New (nobody has it), Acknowledged and In progress (someone owns it), Resolved or False alarm. The tabs are All, New, Being handled and Closed, and the filters narrow the list further. Select several incidents to change their status or Assign to... them in one step.

Open an incident to read its Notes and Status history, add notes, open its recording with Open in Playback, and export its incident PDF or an evidence pack as a zip file. Incidents assigned to you also appear under Assigned to you in the bell.

Video search

Open Video search (Investigate) to find recorded footage by describing it or by showing an example. It needs the Video Search app, switched on and in the licence (Apps), recorded footage, and the AI engine running. If the engine is off, the page says "Search runs on the AI engine, which is not running. Recordings still play in Playback." With no footage indexed it says No footage is indexed yet; turn recording on for a camera and it is searchable within minutes. Two badges say how it works: Runs on this server, so nothing leaves the site, and Searches are logged, so who searched for what is kept. Each person sees only their own searches.

1. Say what you look for. Choose Describe it and type a few words, such as "a person in a red jacket near the loading bay", or choose Example image and pick a picture of up to 25 MiB. A role without access to people has faces covered on the picture, and the picture is deleted after the search.

2. Pick the cameras and the period. A camera that cannot be searched is struck through and says why: switched off, not recording, no picture, or not indexed yet. The period is Last hour, Today, Yesterday, Last 7 days or Custom. Only the first 64 cameras are searched. A line under the page says how far each camera is indexed.

3. Choose how sure a match must be. Show sets Strong only (80% and above), Strong and good (70% and above) or Include possible (60% and above). Sort by Best match or Newest. Matches that are close in time fold into one moment.

4. Read the matches. Each card is a still from the recording with its camera, the time on the site's clock and a score such as "82% match". Point at a card, or focus it, to play its 30-second clip (10 seconds before and 20 after) with the detection boxes drawn on it; with reduced motion, a Play 30 s clip button does this instead. Boxes (key B) switches the boxes. A card says in plain words when there is no still: not recorded yet, no recording at that moment, deleted, or could not load.

5. Open a moment. Selected moment shows the frames 10 seconds before, at and 20 seconds after the match, and a day strip of that camera's other matches. Open in playback opens the recording there. Export 30 s clip needs the export permission. Search with this frame searches for more like it. On a phone it opens as a sheet.

A score for words and a score for a picture are measured differently, so compare scores within one search. Deleting footage also removes it from the index. Searching for vehicles is Later.

Notifications

Notifications reach you while the platform is open in your browser. Their settings are saved in the browser, so each browser you use has its own.

• The bell. It shows the number of new violations since you last closed the list. Choose it to open the list, then an event to open it. The filters are All, Open and High, and Mark all read clears the count. The list holds up to 200 events that arrived while the platform was open; after a lost connection it adds what started meanwhile as unread, without pop-ups. For history, use Events.
• Pop-ups. By default a toast pops up for each new High violation and for account and system notices; Medium and Low violations reach only the bell. A camera raises at most one toast every 30 seconds, and further events in that window are grouped ("3 more on Gate cam"). A toast names the person when they have a name.
• Mute. Choose the bell-off icon on a notification, then Mute for 10 minutes, Mute for 1 hour or Mute for today. A muted camera raises no toasts, sounds or browser notifications. Its events still reach the list. The list shows "muted until" with an Unmute button.
• Browser notifications. Optional. The browser asks for permission when you switch them on. The platform posts them only while the tab is hidden.
• Assigned to you. The bell lists the incidents assigned to you, each linked to its incident.
• System alerts. A camera that goes offline, a stopped vision worker or trainer, lost live updates, and low disk raise an alert. It shows in the System group of the bell, above the events, and a camera offline alert appears within about 70 seconds of its last picture. Choose an alert to open the camera or the System page. Administrators acknowledge or reopen alerts under System > System alerts, and an alert closes itself when the cause is gone.
• Alarms. A check that raises an alarm, such as fire or person down, shows a full-width red banner at the top of every page, above the top bar so the bell stays within reach, and repeats a siren until someone acknowledges it or it clears. Each row reads, for example, "Fire at Dock 2", with the time, an Open button and, for people who may review events, an Acknowledge button. Mute the siren and Unmute the siren control the sound on your screen. A browser plays sound only after one click on the page, so the banner offers Click to enable the alarm sound. A red chip in the top bar counts the open alarms, such as "2 alarms". An alarm that waited while the server was down says "delayed by N min" and never sounds the siren. Alarms reach every open screen. Settings → Notifications sends them elsewhere too (chapter 17).
• Sound. Choose Pop-up with sound for a kind of alert and its pop-up plays a tone: an urgent tone for a High violation and a soft one for other alerts, so a High stands out. Nothing plays until you choose it. A browser can block the sound until you have clicked in the page once.
• Repeat for High violations. Off until you choose Every 30 seconds, Every minute, Every 2 minutes or Every 5 minutes. It needs High violations on Pop-up with sound and the right to acknowledge violations. The urgent tone then plays again, one interval after the page loaded and then once per interval however many violations wait, until someone acknowledges the violation, reviews it as a true violation, or it ends. The count on the bell bumps once for each repeat. A tab that stays hidden for minutes may repeat later than you chose, up to about a minute apart, so keep a control-room tab visible. A muted camera and Do not disturb stop it, and it starts again when Do not disturb ends. When a person cannot use it, the row says why.
• Do not disturb. In Notification settings, the Do not disturb row offers 15 minutes and longer, up to the longest the site allows. While it is on, your pop-ups and tones stop, repeats stop and the bell shows a bell-off icon. The bell still counts every alert, and its list says "Quiet until" with the time and an End now button. Alarms always show the red banner and sound the siren. A wall screen (/monitor?display=wall) keeps its tone and shows no pop-ups, and a display account cannot turn it on: its row says "A wall screen always shows alerts." It follows you on every device, so starting it or ending it in one browser changes the bell in your other browsers within a second. When it ends, a message says how many alerts came in, with an Open bell button. A server without it shows no Do not disturb row.

Under Notification settings in the account menu or the bell, choose for each kind of alert (High violations, Medium violations, Low violations and Account and system notices) whether it is Bell only, Pop-up or Pop-up with sound. Play tries the tone. Every violation still reaches the bell. The defaults are Pop-up for High violations and for notices, and Bell only for Medium and Low. Desktop notices is a separate switch, and the page lists the Muted cameras. Alarms always sound with the siren and a red banner, whatever you choose. Every role has these choices, and they apply to this browser only.

People

The platform gives each person it tracks an id such as P-1A2B3C4D. The same id follows a person from camera to camera, so you can see a journey and everything that person missed. People, names and face pictures are for admins, supervisors and reviewers. A viewer has no People entry in the sidebar, its address answers No access, and the Monitor shows a track number where the person id would be.

The People page

Open People. Switch between cards and a table. Search by name or id.

Filters and sorting

Show filters by All, Named or Unnamed, Seen by 24h, 7d, 30d or All time, and you can filter by camera and by Group.

On site now

On site now lists the people a camera saw within the last few minutes (Settings → Detection, "On site now means seen in the last"); a green dot marks them. A person who was enrolled but never seen reads Not seen yet.

Faces and readability

A short notice on the page reminds you that faces are biometric data. Sort by Last seen, Most violations or Name. People under 100 px tall are counted but not listed. Face readability shows, for each camera over the last 7 days, how often a face was big and clear enough to recognise; a low figure means recognition will rarely work on that camera.

A person's card

A card shows the sightings, the cameras and the violations of one person. Open a person for their activity timeline, the cameras they were seen on, the PPE they missed most and links to their events.

Naming people

Reviewers, supervisors and administrators can name a person: choose Name this person on their page or menu, type a name and add optional notes. The name belongs to the id. It shows next to the id on Events, in notifications and in the Summary, and it returns whenever the cameras recognise that id again. With face recognition on (chapter 8), a named person is also recognised by face on later days.

Enrolling people and groups

Administrators and supervisors can add people before a camera has seen them. Choose Enrol person, then Enrol a person from photos opens: add one to five clear photos, each with one face (JPEG or PNG, 10 MB at most), a name, and optionally a group. The dialog checks the count, size and type before it sends, warns when face recognition is off, and shows the server's answer for each photo, so you can try better ones. A person keeps at most ten enrolment photos. Add photos on a named person's menu or page adds more (the title reads "Add photos for" and the name). A failed enrolment leaves nothing behind.

Groups (administrators) opens a dialog to add, rename and remove groups such as staff, contractors and visitors. A person has one group and can be on the Watchlist, with an optional reason; set both when you name them. The page of a person shows their group, the watchlist reason and how many enrolment photos they have. People who can see people can see groups.

From a person's menu, Delete face data… removes their enrolment photos, face templates and face picture; the name, the sightings and the events stay, and they are no longer recognised by face until you enrol them again. Delete person… removes that and also the name, group and watchlist entry; sightings and events stay under the bare id. Both ask first. Someone who is forgotten and then seen live is not learnt again.

Merging two ids

When one person ends up under two ids, open one of them and choose Same person as…. Pick the other id, then choose which id to keep. The other id's sightings, events and enrolment photos move to it, and the cameras treat both as one person from then on. The dialog is titled Merge into one person?. Naming and merging are recorded in the audit log.

Visits

A visit is one stay of a person in a zone. The Visits tab on a person's page lists their visits day by day, and each sighting names its zone. A stay is one sighting, however many frames it spans. When a person has been out of sight in a zone for longer than Time out of sight before a new visit (Settings → Re-identification), seeing them there again starts a new visit. The default is 60 seconds. Merging two ids moves their visits to the id you keep. Visits are kept for 90 days by default. A visit whose camera stopped watching near its start or end reads "at least" before its duration, and the page marks its figures as estimates.

The Visits page lists every visit, newest first, with a summary of the visits, their average duration and the longest one for the period and filters you chose. Filter by period, area, zone, person and duration, or show only visits missing PPE or only people still in a zone. Choosing a visit opens its side panel. The list exports as CSV (Export N visits (CSV)), with a column for the area and a note of the site's time zone. Only roles that can see people can open it.

Limits you should know

• Matching uses appearance: clothing, build and colour. People in identical uniforms can be merged into one id, and a person who changes clothes can look new.
• Face recognition helps only for named people, and only when the face is visible.
• An id is a strong lead, not proof of who someone is.

Administrators can tune matching under Settings → Re-identification. If one person gets many ids, raise reid.maxGapS or lower reid.threshold. If different people merge, raise reid.threshold and switch off crossCameraOverlap. The defaults are a threshold of 0.72 and a gap of 180 seconds.

Overview and Reports

Overview

Overview is the home page for administrators and viewers. Under the title it says how many cameras there are and how many are reporting, with the time and its time zone.

Tiles

The tiles are Open violations (events still in progress), Awaiting review (events nobody has reviewed yet), the Active model and System (the vision service's state, its provider and its load, with a link to the System page). Reviewers and above also see Pending samples.

Needs attention

Below the tiles, Needs attention lists the worst open violations nobody has reviewed first, without Test bench runs, then cameras that stopped sending pictures (one that never sent any reads "No video yet"), events awaiting review, and a vision service or PPE model that is not working. It never shows a tick while a camera is unchecked. Each row opens the page where it is dealt with.

Trend and cameras

Violations charts the last 24 hours by hour, or the last 7 days by day, against the period before. Cameras shows each camera with its coverage over the last 24 hours. The six latest violations follow.

Clock changes and coverage

On a day the clocks change, today starts at the real local midnight and the 7-day trend gives a 23-hour or 25-hour day its own bar. Violations and the Cameras strip carry the coverage and rate lines described under How the numbers are counted.

Area select

The Area select in the header makes the counts, the trend and the latest violations about one area; Needs attention and the coverage figures still cover the whole site, and the page says so.

A new site shows Set up this site above the tiles: a short list of steps from replacing the install password to adding the people who will use the platform, then requiring two-step sign-in for administrators. Each step ticks itself when the work is really done, and its button opens the page where it is done. To leave a step for later, choose Skip and say why (5 to 200 characters), so the next administrator knows; Undo brings it back. Someone whose role cannot do a step sees why, in words, beside it. When every step is done or skipped the list says Set-up is finished. and hides itself a week later, or at once with Hide it now. An administrator can switch the list off under Settings → Site, Set-up list on Overview.

The first time an administrator opens Overview on a new site, a short welcome asks where the site is. Its time zone sets the clock for reports, rule schedules and every time shown, and the welcome says which privacy profile applies. It shows once for each administrator. The list also asks for the site's privacy country, two-step sign-in for administrators and the nightly backup. Two-step must be on for your own account before it can be required for administrators, and a site that does not want it skips that step with a reason.

How the numbers are counted

Every figure on Overview and Monitor follows the same definitions.

• Time in compliance is the share of judged person-time that was compliant. Every second a person is present is exactly one of compliant, violating, could not decide (too small, too new, or no PPE model) or out of scope. The seconds before a violation is declared count as violating. The figure always shows a best and a worst case, with all the undecided time one way and then the other; when less than 70% of the time was judged, only those two bounds are shown.
• Coverage is the camera time observed over the camera time expected inside the working hours set under Settings → Counting & coverage (no windows means all hours). Under 90% a figure says "partial"; under 50% it shows "—" and names the camera. If a camera is online but the server checks it less often than it is set to, the page says so ("The server checked 30% of what this camera is set to. Lower its checks per second or add capacity.").
• Violations show "N per 100 judged person-hours" once there are 10 judged person-hours. An arrow compares with the previous period on the cameras observed in both ("compared on 12 of 14 cameras"). Events reviewed as false alarms leave the count and are shown beside it ("plus 4 reviewed as false alarms"). Test-bench runs are in no figure.
• No data is not zero. A period with nothing observed shows "—" and says why.
• Estimates. "People seen (estimate)", "Compliant so far" and "Non-compliant so far" in the Monitor's session view count confirmed tracks since the stream started, and can count a person again if they leave and return. A person is counted only after being tracked for a short while at a usable size. Every people and visit figure carries How counted, which says what one unit is and that a person may count more than once.
• Silent cameras. A camera that stops reporting for 3 minutes has its open violations closed at its last heartbeat; such an event has no duration for statistics.

Reports

Reports (Investigate → Reports) replaced the Analytics page; the old Analytics address still opens it. Every role that can read Events can open it. The Compliance tab carries no names, faces or person ids. Reports has up to five tabs. Compliance summarises a period for the whole site, Visits is the visit log, Incidents is the incident log and Vehicles reports crossings by type and hour on each line and every wrong-way alert (chapter 6). Validation, for administrators, is the model validation report (chapter 16). The Visits tab lists people's visits, so it needs access to People. A camera health report is In progress.

Available

• Period. Pick Today, Yesterday, Last 7 days, Last 30 days, This month or Last month, or choose Custom and set the two dates (at most 366 days). The site's time zone is shown beside the period ("Times in Asia/Colombo"): every day, hour and time in a report follows it. Set it, and the work shifts, under Settings → Site.
• Filters. Area, Camera, PPE type and Shift (the Shift filter appears once a shift exists). The camera filter narrows everything. A PPE type or a shift narrows the tables and the hour grid only; the five tiles still cover every PPE type and shift, and the page says so. Clear all keeps the period.
• The tiles. Time in compliance (with the range it lies in, or only the range when too little time could be judged), Violations (with the rate per 100 judged person-hours), Ids seen (estimate) (a person who leaves and returns can count again), the most-missed PPE and the peak hour, each against the previous period of the same length. How counted on a tile opens its definition, the same one Overview uses. A sentence under the filters says how much of the expected camera time was observed.
• The panels. Violations per day against the days before, Reviewed so far (how many events a person has looked at; it is not an accuracy figure), By PPE type, When violations happen (hour by weekday; press a cell for its count), Top zones, By area (each area's violations and time in compliance) and By camera with how long each camera was up. By check is a table of events, alarms and the events confirmed in review for each check, counts rather than an accuracy figure. Ignored at fixed items shows how many detections each camera's fixed items ignored. Both follow the filters and appear in the PDF and CSV too. Where to look first ranks the zones, or with Shift the shifts, where problems cluster; it never ranks people.
• Download. PDF report is a branded A4 file with the licence's logo and who made it and when; CSV data opens correctly in Excel, one row per day, camera and PPE type. Both carry the period and filters you have on screen and are named for them. Camera, zone and person names in Sinhala or Tamil print in report and incident PDFs; the screens are in English only. A person can download 30 files a minute, and each download is in the audit log.
• Incident and visit logs. The Incidents tab downloads the incident log as CSV, JSON or PDF. The Visits tab downloads the visit log as a landscape PDF or as CSV. The CSV files carry an Area column. Zones that share a name show their camera, so they can be told apart.
• Customize. Choose which panels show, in what order, whether the marks for the period before appear, the row spacing (Comfortable or Compact), and the area, PPE type and shift a report opens on. It is kept for you and Reset to default undoes it.

Use it to spot a shift or an hour with more misses, the item people forget most, and a camera that raises many false alarms. A camera with many false alarms needs attention: look at its view, its zones and its rules, and feed it training data.

Test bench

The Test bench checks PPE on an image or a stored video without a live camera. Use it to see how the active PPE model performs on your own footage, and to compare views side by side. Administrators and supervisors can use it. It is under Improve → Test bench.

A pill at the top names the active PPE model. It warns when no PPE model is active or the vision worker is offline.

Set up a run

The setup card holds one run, A. Choose a Person model and a PPE model for it. Only the test bench uses them; the live cameras keep the active model. Turn on Compare to add run B, which checks each file a second way, so two models or modes can be compared. Required PPE sets the items that count. Required checks names the safety checks a video run must exercise, such as Fire or Phone use; it lists only checks whose app is on, and says "No safety checks are available." when there are none. The run card shows them, and a picture you drop shows those checks' boxes, so you can try a check before you switch it on for a camera. Second opinion makes another model look again at each person for items the first missed. On a picture its finds are drawn dashed, and the card says how many verdicts it would change. On a video it is a third PPE model: it is shown, never counted, and the finished run says how many people it found extra items on. A looping video uses the live models only, so it takes no third model.

Add files by dropping them on the page, choosing them from your computer, picking a stored video from the library, or pasting an image. On an empty bench, Load the sample runs runs three sample photos and one short video through your models, so you can see a result before you add your own files. Every result names the models and the mode that made it, and a required item the model cannot detect is reported as not checked, not as missing. A person who could not be checked for every required item reads as not decided, never as compliant.

An image is one frame. There is no tracking and no time evidence, so a result is one guess. Do not read a single image as proof of how a live camera will behave.

Results

A photo or video shows a summary with the totals, and a card per scene. Its tabs are People (an OK tag or the items each person misses), Detections and Info. A colour key explains the boxes, and Layers chooses what is drawn. Under a finished video, People in view now follows the playhead, and the run says how long processing took, in GPU seconds on a GPU or CPU seconds otherwise. Side by side and Slider compare A and B, and the Differences panel lists where they disagree. PPE boxes take the colour of their PPE type (chapter 7).

Drop a video (mp4, mkv, mov, avi or webm), or choose Add from library for one you uploaded before. Before you start, set Required PPE, Checks per second, Loop and Record clips. A video run goes through the full pipeline: tracking, zones, time-based rules and real events. Each run has a timeline, its tracks and a report, and its results download as CSV. A run says how many of its frames were actually checked, such as "Checked 1,180 of 1,200 frames (98 %)", so a run that skipped frames is easy to spot. A run raises real events that you can open and review like any other. Removing a run deletes its events and keeps the uploaded video.

Download zip exports the results as one zip file. A zip that is too large is refused, so export fewer runs at a time. On a phone the setup is a read-only summary, and the results switch between A, B, Slider and Diff.

Replay with boxes

Open a video run to replay the original video with the boxes the run found drawn over it, with Boxes and Labels switches and a timeline strip with a tick at each violation. After Reanalyze, Before and After replay the run as it was and as it is now. Reanalyze replaces only the saved boxes and keeps the video. If a run's original video is no longer stored, the page says so.

Fast, Accurate or Max

The mode is set per run in the setup card. The GPU or CPU chip in the top bar holds the same three modes. Fast is the live cameras' setting and finds the least. Accurate doubles the resolution, adds extra passes, runs slower and finds more. Max also searches tiles for far-away people, and on a 1080p picture it takes about 5 times as long as Accurate (measured on a CPU server).

Improving accuracy

The platform gets better when you teach it what your site looks like. This chapter covers the models and the loop that improves them.

Models

Only administrators see Models. Two models work together: a person model that finds people, and a PPE model that finds the items. The panels at the top show which is active.

• Model library. A catalog of ready-made detectors. Pick a family and a size (the Recommended one is a good start), choose Download, then Activate once it shows Installed. The platform reads the model's classes and suggests a class map. Ultralytics models run in every edition. Filters above the library narrow it to All, PPE, Person and, with the Safety Checks app on, Hazards, Behaviours, Pose and Objects. A link from the Safety checks tab opens the library for that check. All models opens every model that could be added, 664 for 62 jobs: search them, read what each does, and shortlist one per job. The shortlist is kept on the server, so every administrator sees the same one. All models installs nothing itself; Install from the Model library opens the Models page at that model.
• Default person detector. A new site downloads the pinned D-FINE small person detector, under the Apache-2.0 licence, which runs on ONNX Runtime; a site that already runs a person model keeps it. If the GPU cannot run a downloaded person model, it runs on the processor. If the active person model cannot load at all, the System page says so and the built-in detector finds people.
• Running now. A table lists the models the AI engine holds in memory, with the memory, speed and cameras of each (Running now, chapter 6). On Intel hardware, detection can run through OpenVINO, which was faster than the standard engine on the same processor in our test, and is picked automatically once installed.
• Qualification card. Each model shows its licence and commercial use, whether its checksum was verified, whether it loads on this PC, its speed and its accuracy. Choose Run benchmark or Re-run benchmark on the card. Models installed from the library are benchmarked automatically. An operator can switch on a signed online model feed and the discovery of new releases, which show as unverified. Both are off by default.
• Hazard, behaviour and pose models. With the Safety Checks app on, the Models page also lists them. Several of each can be active at once, so activating one skips the golden gate and its button reads Deactivate; person and PPE keep one active each. The library carries hazard models for fire and smoke, forklifts and construction hazards, official pose models, and an experimental cigarette detector for the Smoking check. Each is pinned to a release and checked by its checksum on install, and the hazard and cigarette ones carry AGPL-3.0 terms. RTMPose, a pose model that runs on every installation and is quick on a processor, lets Person down use its pose method without the PyTorch engine; other pose models need it. A second fire and smoke model, trained on outdoor and aerial scenes, suits yards and open sites and runs on the Ultralytics engine that every release build includes.
• Model kinds and fit. Each library card says what its model hands back: Boxes, Outlines, Rotated boxes, Body points, Whole-picture label, Outlines from a prompt, Depth map or Text. Each size shows a fit chip: Runs on the CPU, Faster on a GPU or Needs a GPU. On a server with no usable GPU, a size that needs one warns you on its card. Objects in the task filter lists the object models.
• Object models. Installed object models sit in their own section, several can be active at once, and each has a Run on live cameras switch and a Follow each object switch. A heavy model starts switched off, with the note "Heavy for this server: switched off until you turn it on"; turn it on when the server can carry it. A model that takes typed words (YOLOE and YOLO-World) offers Type what to find: list up to 40 names, one per line, and the platform makes a candidate copy that looks only for those things, ready to test before you use it. The Object seen rule in Rules raises an event for the object you pick, from any object your installed object models can find, and stays idle until you pick what to look for.
• Bake a model. This menu at the top of the page offers Person down (YOLOE-26s), PPE (YOLOE-11L) and Hygiene & food safety (YOLOE-11L), each while its app is on. The trainer turns text prompts into a candidate model within a few minutes, which you activate yourself. It needs the trainer's open-vocabulary add-on, installed by hand.
• New releases (unverified). An operator can switch on discovery of new Ultralytics person detectors. They are listed here but never install in one click: you confirm that they have no published checksum first (Install as candidate).
• Add from URL. Downloads a YOLO .pt or .onnx file from an address you give. You can add a licence note. A release build installs only models with a published checksum, so it turns Add from URL and Install as candidate off and says why: "This build installs only models with a published checksum."
• Import model. Registers a .pt or .onnx file from your computer. Give it a name and a task, person or PPE.
• Candidates. Each model has a status: candidate, validated, rejected, active or archived. When it has been scored, its card shows Golden-set score, Score with tight boxes, Right when it reports and Real items found, each with what counts as good: a golden-set score above 70% is good, and under 50% misses a lot. Mark as… sets a status, Make ONNX copy adds an ONNX copy of a .pt model as a new candidate, and Delete removes an inactive model's file. Built-in models have cards too.
• Evaluate and compare. Evaluate on the golden set scores any person or PPE model on the golden frames, with ranges, misses and a confusion matrix. Compare with active shows two models class by class and says plainly when the new one is worse. Validation report (PDF) prints the model validation report, which Reports → Validation also gives as PDF or CSV.

Class map

A PPE model names its classes its own way. The class map says what each class means for the platform. Choose Class map on a model, then for each class pick the PPE type it is evidence for. Tick absence for classes such as NO-Hardhat, which prove that an item is missing and never count as the item being present. Leave a class on ignore when it is not PPE. Auto-suggest fills the map from keywords. A saved map reaches vision within 15 seconds, with no restart. If the active model maps to no PPE type, the System page shows the ppe_model_unmapped warning.

Activate, force and roll back

Activating a model is deliberate.

The golden-set gate

The platform scores every candidate on the site's test images, the golden set, and compares its golden-set score with Minimum score to go live in Settings → Training. A new site starts at 85%; a site that saved its own value keeps it.

Item-by-item gates

Activate also compares the new model with the live one, item by item. An item that scores more than Largest drop allowed on one item (2 points) lower than the live model, or under Minimum score for each item (50%), stops it too, and the page names the item.

When a gate is missed

A model that reaches every gate activates after a confirmation. A model that has no score or misses a gate opens Golden gate not passed, where you type a Reason for overriding of at least 10 characters and choose Force activate. The reason is kept in the audit log. Imported and downloaded models have no score until a training job or Evaluate on the golden set scores them, so an unscored first activation needs this.

Re-inspect

Re-inspect on an imported ONNX model reads its class names again.

Activation archives the model that was active, and vision reloads within 15 seconds. Roll back on the active model restores the previous one. If a new PPE model cannot load, for example because a training job holds the GPU memory, vision raises ppe_model_load_failed. PPE cameras then run without PPE checks until it loads, and vision retries every 5 minutes.

The improvement loop

1. The platform collects frames worth a look. It queues frames where a PPE item sits in the uncertain band, where person confidence is low, or where the scene looks new. Marking an event Not a violation also queues its frame. The limits sit under Settings → Mining.

2. You review them. Open Datasets → Review queue. Fix the boxes and classes, then choose Approve, Reject or Add to golden.

3. You freeze a version. Open Datasets → Versions and choose Create version. A version is an unchangeable snapshot of the approved samples. Its training and test images are split by camera and hour, and a picture taken within 10 minutes of a test image or an accuracy check on the same camera is left out of training; the version's notes say how many.

4. The platform trains. Start a job under Training, or let automatic training do it. A PPE job fine-tunes the active model by default. The Detector choice in New training job is YOLO (Ultralytics), AGPL-3.0, or RF-DETR, Apache-2.0, so the result may be sold with the product; the licence shows under it. RF-DETR trains from its own pretrained weights, has a Size instead of a base model, and is In progress: its training add-on is not installed by default and a first full run is still to be confirmed.

5. You check the result. The finished job is a candidate with golden-set scores on the Models page.

6. An administrator activates it, through the golden gate and the item-by-item check. Roll back if live results get worse.

Datasets

Datasets → Review queue shows counts of pending, approved, rejected and golden samples, a list of pending samples on the left and a canvas in the middle. A sample's source is uncertain, novel, event review or manual, and the list shows its priority.

Draw boxes

Pick the class for new boxes, draw boxes on the canvas, zoom in on small items, and drag or delete the wrong ones. Reset to model labels brings back the model's own boxes.

Suggest boxes

Suggest boxes draws dashed boxes from the Active PPE model, an Open-vocabulary model or the Built-in labeller, for you to accept or discard.

Filter and bulk actions

Filter the queue by Status, Camera, Class and Date, or select several frames to Approve, Reject or Delete them in one step.

Labels

The Labels card imports YOLO labels (.txt), COCO labels (.json) and pictures or a zip, up to 256 MB at a time, and exports this frame, the selected frames or a frozen version as YOLO or COCO.

Upload and approve

Upload image adds your own picture, optionally for a camera. Approve and Add to golden need at least one box. Otherwise reject the sample.

Golden samples are permanent test data. Every candidate is scored on them, and they are never trained on. Add a spread of cameras, hours and lighting to them: a golden set that is small or stale makes the check weak. Versions lists each version with its status (building, ready or failed), samples and classes. Creating a version needs the Supervisor or Admin role. Train starts a job from a ready version.

Training

Open Training (Admin and Supervisor). Use New training job to start one: choose the task, a base model (the default is to fine-tune the active model; it must be a .pt model, and an ONNX or engine model is refused at once with a message asking for a .pt model), a dataset version and, if you wish, epochs, image size, batch size and learning rate. Export ONNX also saves an ONNX copy. Leave lr0 empty and each detector uses its own learning rate (0.01 for YOLO, 0.0001 for RF-DETR). A band at the top says what the trainer is doing: the running job with its epoch, a progress bar and the time left, or a job waiting for the idle hours, or idle. It holds Pause, Resume and Stop. The job panel below shows the job’s Curves or Log. A cancelled job cannot be resumed. Past runs lists every PPE model with its golden-set score against the gate, and an administrator can Activate one, Activate anyway with a written reason, or Roll back. The Scheduler panel says who holds training and why, with the idle hours and the live camera load of the last six hours. All jobs lists every job with its status, scores and log.

• Idle windows. Jobs run only in idle windows, by default every day from 22:00 to 06:00 on the site's clock, and training pauses when vision is busy (load above 0.7). A job also waits until Free memory needed to start training (2 GB) is free and, on a graphics card, takes at most Largest share of graphics memory for training (60%), so live detection is never starved; the job says what it is waiting for. Pausing, resuming and cancelling happen at the end of an epoch. Tick Ignore idle windows on a job to run it now; such a job has no Pause. If someone paused training, Return to automatic scheduling hands control back.
• Automatic training. Switch it on at the bottom of the Training page (administrators) and the platform creates a version and a job when 200 new approved samples (minNewSamples) have built up. It is off by default.
• How much data. The first fine-tune from a ready-made model rebuilds the detection head. Plan on a few hundred labelled frames per class.

Accuracy checks

Open Improve → Accuracy checks (Admin, Supervisor and Reviewer) to measure how often detection is right on each camera. The answers never change an event.

Three figures sit at the top of the page. This week shows how many of this week's questions you have answered, with the minutes left. To settle counts the questions that two people answered differently and that wait to be settled. Sets due counts the cameras that have no check set yet, or whose check set is out of date. Each figure opens its tab. See the results opens the Accuracy report on Reports.

• This week. A blind set of short moments from the cameras, shown without what the system decided. Say what you saw, and the next question appears.
• Disagreements. The questions that two people answered differently.
• Camera sets. Each camera can have its own check set, so it is measured on its own pictures. An administrator or supervisor chooses Collect a check set, which gathers it over four working days. A reviewer then chooses Mark moments, fixes the boxes and answers for each person.
• Staged walks. On Camera sets, Start a walk measures how often a missing item is caught. Someone walks past the camera without one item at a time, near, at middle distance and far, and the page offers Start this pass and End this pass for each pass, then Finish the walk. Walk without an item only during a stoppage or under supervision: it may break a site rule. It needs permission to change cameras.
• Short clips for tracking. Collect clips gathers eight short clips for the site, and Label a clip lets a reviewer give each person the same label in every frame. The answers measure how often the platform loses or swaps people, and they tune how pieces of a path are joined.

Weekly accuracy check (on by default) and Questions per week (100 by default; every camera gets at least 4) are under Settings → Counting & coverage. The Storage page lists these pictures on their own line, Accuracy check evidence. Check-set pictures are never offered for clean-up, and weekly check pictures are removed after 90 days.

Administration

Users

Open Users (Admin only). The table shows each person's name, email, role, status, creation date and last sign-in. Add user asks for an email, a name, a password of at least 8 characters and a role. Edit changes the name, the role and the password (leave it blank to keep it), and its Disabled switch signs the person out and blocks them. You cannot change an email. Delete ends the person's sessions at once, and the audit entries stay. You cannot delete your own account.

• End someone's Do not disturb. A person who has Do not disturb on shows a Quiet until label on their row with an End button. It returns their pop-ups and tones on every screen at once and is written to the audit log.
• Approve or reject a request. A new request rings the bell with a Review button, and the Users entry in the sidebar shows how many are waiting. The Requests tab lists each one with the name, email, when it was sent and from which network address, and the person's note. Pick a role and choose Approve, or choose Reject with an optional reason that only the audit log keeps. Refused requests are deleted after 30 days.
• Invite someone. Choose Invite, enter a name, an email and a role, and choose Create invitation. Copy the link and send it yourself. It works once, for 3 days by default, and is shown only once.
• Follow an invitation. Until the person has chosen a password, their row reads Invited, expires and a date. Choose Resend link for a new link (the earlier one stops working), or Revoke link in the row's menu to end it; the account stays and the row reads Link revoked. Someone waiting for approval reads Pending approval.
• Display accounts. The Display account switch makes a sign-in for a wall screen. Its role is locked to Viewer, the account shows a Display badge, and it stays signed in for 30 days at a time while the screen is on.
• Reset a password. People who chose Forgot password? also wait on the Requests tab. Choose Create reset link (the request leaves the queue once you have copied the link; closing early asks you to keep the dialog open or close anyway) for a one-time link that works for 1 hour by default, or Set temporary password to type one or Generate one you can read out. Either way the person is signed out everywhere and chooses their own password. Reset… on a row in All users does the same for anyone else.

Apps

Open Apps (Admin group; administrators only) to choose which apps this site uses. Each built app has a card with its switch, what it adds to the menu and what it needs. People who cannot change apps see the same cards with their role named. The apps you can use are CCTV & Recording, PPE Compliance, Security, People & Visitors, Safety Checks, Video Search, Model Training, People Counting, Vehicles & Plates, Conveyor Counting, Retail & Area Analytics, Sound & Voice, Time & Attendance and Visitor Kiosk. CCTV & Recording is the base app and is always on. People Counting adds counting lines, the Counting page (chapter 6) and the Occupancy limit check. Sound & Voice adds a Sound page (Operate). Time & Attendance needs People & Visitors and adds an Attendance page (Investigate). Visitor Kiosk needs People & Visitors and PPE Compliance and adds Expected visitors (Operate). Conveyor Counting adds a Conveyor page (Operate), Belt counting lines on the camera page and an Operator panel section in Settings (chapter 6); a site that only counts a belt can switch every other app off. Two apps do only part of their job so far:

• Vehicles & Plates adds a Vehicles page (Operate), a Vehicles section in Settings and the Wrong way and Stayed too long in a bay checks. The page counts vehicles crossing a line, in and out, and shows Parking bays: a grid of the bays drawn on a camera, each taken, free or reserved. Choose Set up lines and bays and the Parking bays tab to draw a Row of bays (one shape split into equal bays) or a single Bay; each bay has a Number, can be Reserved (Reserved for names who) and can have a Longest stay, from 5 minutes, after which one low Stayed too long in a bay event is raised. Click a bay to set it Taken or Free by hand; it shows "Set by hand at 10:12" and holds until the camera sees a different state. Setting a bay by hand needs permission to review events. Reports → Vehicles adds the share of bays taken by hour, the average stay and the ten longest stays to its PDF. It also reads number plates at a gate line (chapter 6). Vehicle speed is Planned.
• Retail & Area Analytics adds the Queue point check. Heatmaps and time in each zone are In progress.

Time & Attendance is off until the site sets it up. It clocks people in and out from recognised faces; Settings → Attendance holds its options.

• Switching on and off. Switching an app off asks first and lists what leaves the menu. A message then offers Undo for a few seconds. An app the licence does not include stays locked.
• What stops and what stays. A switched-off app stops its live analysis, its new events and its models. Recordings, events, rules, people records and history stay, and switching the app on again picks up where it was.
• Running now. A panel at the top shows which AI models the vision service holds in memory. It unloads a model that no app on needs, so switching an app off frees the GPU. A stale heartbeat reads offline.
• Dependencies. An app that needs another says so on its card, "Needs People & Visitors, which is off." for example, and one click switches both on.
• PPE Compliance is the same switch as Settings → Detection → PPE checking.
• Overview. With no AI app on, the Overview hides its violation, review and model tiles and the trend, and Needs attention lists no events. Settings lists only the sections of apps that are on.

Settings

Open Settings (Admin only). Every setting has a plain name, one line on what it changes, its default and its range, and a way back to the default once changed. Less used ones sit under Advanced, which opens by itself when one of them is changed or wrong. A setting that cannot be saved says what is wrong, in plain words, beside its field. Choose Save changes in the bar that appears at the bottom, or Discard. Changes reach the cameras' checks within about 15 seconds and training from its next job; nothing needs a restart. If someone else saves while you are editing, your draft stays, and you are asked only when you both changed the same section. Four settings draw a small picture of what they mean: where a PPE item counts as missing, checking or worn; the free-space limits against today's free space; the hours of the week when training may run; and which camera can follow which.

Setup backup

An administrator who can change settings and manage users sees Settings → Setup backup. Download setup backup saves everything this site is set up with in one file, locked with a file password you choose and typed again to confirm. The file holds the settings, the accounts with their password hashes, the cameras with their zones, rules, lines, bays and floor maps, the models and their weight files, the notification and mail set-up, and the enrolment photos of named people. Camera, mail and two-step secrets are kept too, so a restore on a new server opens them; where no secret is kept, you type camera and mail passwords again, and two-step sign-in is set up again. Events, recordings, counts, the audit trail and the licence are not in it. Keep the file like a password.

Restore from a file puts a file back. Choose the file, type its file password and choose Check file. The page shows what the file holds beside what is here, and warns about a camera here that the file lacks (it stays, with its events), about your own account when it is not in the file (you are signed out after the restore), and about unsaved changes on other Settings sections. Type your account password and choose Restore. Before anything changes, the current set-up is saved as an undo file: Download undo file brings it back, and it opens with the account password used for that restore. A restore replaces the set-up but never deletes history: events, recordings, counts and the audit trail stay.

Notification channels and routes

Open Settings → Notifications to send alarms beyond the app.

Email

Email goes through one Mail server panel. Pick a Provider: Own mail server, Microsoft 365, Google or Other provider. It fills in the usual server, port and Security: STARTTLS (usually port 587), TLS from the start (usually port 465) or None, for a server inside your network. Fill in the account and choose Send a test: it sends one short email before you save, and a failure names the step that failed. The test arrives as formatted text with a plain-text copy, like an alarm email.

Channels

Then add a channel of one type: Webhook (with an optional HMAC signing key), Email, SMS, WhatsApp, HTTP siren / relay or Fire-panel watchdog. An email channel sends to a group: tick Roles and People, or type Other addresses, and each person gets their own email. Each channel can be tested (the watchdog has no test) and switched off.

Secrets

Its secret (the mail password, an auth token, a signing key or an authorization header value) is stored encrypted and never shown again. The API needs an APP_SECRET of at least 32 characters for this, and a new APP_SECRET makes stored secrets unreadable.

Test bench

Test bench runs never notify.

• Routes choose which checks, cameras, zones and severities go to which channels, or only events that raise an alarm. Escalate unacknowledged alarms sends an alarm to more channels when nobody acknowledges it after Escalate after (at least 30 seconds). Quiet hours hold messages in a window on the site's clock (22:00 to 06:00 unless you change it) on the days you pick; a window past midnight belongs to the day it starts. Alarms still go.
• Delivery log lists each message as sent, queued, skipped or failed, and refreshes every 10 seconds. Show older pages back, and Retry sends a failed message again. An email that cannot be sent is tried again for up to an hour (Keep trying for, 60 minutes) before it counts as failed, and a late alarm says how late.
• Email events (on the same page) lists each system email, such as Camera offline: the Least severity, Who gets it and the Wait before it is sent, so a short blip sends nothing. When the problem clears, one back-to-normal email follows. Many notices of one kind in a burst fold into a single summary. The table is dimmed until a mail server is set, and a row that says Nobody gets this email has no recipients. Every automatic email ends with a Stop these emails link, and each person can also choose Email preferences in their user menu. Alarms and sign-in emails cannot be turned off.
• HTTP siren / relay switches an output on with the alarm and sends the off request when someone acknowledges it, or after a set number of seconds. If the server stops responding, detection switches on the outputs that allow it itself and keeps the events to send later; a replayed alarm switches no physical output.
• Fire-panel watchdog pings a fire panel or monitor every 30 seconds while detection is healthy and stops when it is not, so the panel raises a fault.

The System page

Open System (every role). It refreshes every few seconds.

• The header shows the build and two pills for the database and Redis.
• Vision shows the heartbeat (it turns red when stale), the provider and device, a load meter, the active models, GPU memory and utilisation, and its warnings in plain words, such as a missing PPE model or a GPU that is not available. It also shows go2rtc and the number of streams.
• Licence. Administrators see how the AI engine reads the licence. A difference from the server's reading shows in red with the engine's reason.
• Versions. Each part shows its own version as x.y.z.
• Security check (administrators only) lists five go-live checks: developer tools, the service password, the key that seals stored passwords, sign-in cookies, and access requests from the internet. Each row reads OK, Look at this or Fix before going live, with what to set. It shows no value.
• System alerts lists the open alerts first, then recent ones. An alert is raised within 90 seconds and also reaches the bell. Administrators choose Acknowledge to show someone has seen it, or Reopen on an acknowledged alert to bring it back. A site with helper servers, which take some of the cameras, gets an alert when one stops answering, needs an update, has its clock off or loses pictures, and a critical one when some cameras have no server with room. Each clears itself. Each open alert also says in one line what to do about it. Two weeks before the face data of people not seen for a year is deleted, an alert says how many people will be deleted and on which day.
• Servers. With a helper server added, the Servers panel lists the main server and every helper with its state and how busy it is, and Recent moves lists the cameras that moved between servers.
• Capacity. Run a load test plays a Test bench video on test cameras for the minutes you set, then says how many cameras fit. It adds load to the live system, so run it out of hours. The plan has three sliders: cameras, Checks per second and Checks on each camera (1 to 4, added in the order PPE, Security, Safety, Counting), and its estimate scales with each check's cost. The folded Cost of each check card lists what each check adds to a picture on this PC, each tagged Measured once the server has measured it and Typical figure until then.
• Acceptance. The nine items of the contract, each with the date of its last run and what it is tested with. The acceptance script records the performance item here; an administrator presses Sign off on an item checked by hand, and Withdraw takes it back. The passed items fold away.
• Backups shows when the last good backup finished, its size, what it holds and where it is, then the newest runs with their result and time taken. When no good backup has finished in 36 hours, or the last one failed, a warning opens in the bell and on this panel, and it clears after the next good backup. A restore test (scripts/restore-check.sh, run monthly by a timer or by hand, see docs/BACKUP.md) restores the newest backup into a scratch database and checks it. The panel shows when the last test ran and whether it passed, and a warning opens when no test has passed in the days set under Settings → General & branding → Backups. Administrators change or turn off the warning there. A server that uses the built-in database has no nightly backup, so the panel says so instead. Backups made before this update are not listed.
• Download support file (administrators, top of the page) makes one zip for your system provider's support team: versions, health, settings, recent warnings, camera and alert-channel status, licence, updates, backups and database row counts. Passwords, keys, email addresses, faces, pictures and people's names are taken out; IP addresses stay. Nothing is sent: you choose who gets it.
• Training shows the trainer, the current job, the device and the control state.
• Storage shows disk used and free, a state (Normal, Saving space or Critical), the forecast, the three kinds of file that use the most space, and a link to the Storage page. It also shows host CPU and memory. A meter turns amber at 70 percent and red at 90 percent.
• Cameras lists each camera's state, what the camera sends per second, checks per second, latency, people in view, last frame, recording state (recording, stopped, error, disabled or paused) and last error.
• Slowed checks. When the computer stays overloaded, loitering, phone, smoking and forklift-proximity checks run at half rate, one by one, and System lists them. The far look of Find far people and vehicles can also drop to half rate, listed as "Far people and vehicles, half rate". Fire, smoke and person-down checks are never slowed. If that is not enough, the header shows a Not enough compute chip that links to System, and System warns "Not enough compute for the enabled checks". Fewer cameras, a lower inference rate, smaller models or a GPU will help.
• Audit log (Admin and Supervisor) lists the last 100 actions: time, user, action, target and details.

Helper servers

Open Settings → Helper servers (administrators) to let other computers of yours share the work when this server is full. Pick the site's country in Settings → Privacy first: helper servers follow its rules for sending footage to other computers.

1. Choose Add a helper server. Name the computer, say whether it is On this site or Somewhere else (then pick its Country), and tick This computer belongs to my organisation. Choose Add and get a code.

2. The window shows a join code, how long it works, This server's address (pick one the other computer can reach) and a command for Linux, macOS and Windows, each with Copy. Run it on the other computer, where the platform is already installed. It joins within a minute.

• The card of a helper. It shows the state, how busy the helper is, its cameras, version, link time and when it was last heard. Make a new code works while it waits to join. Pause stops giving it cameras, Empty moves its cameras to other servers, and Remove cuts it off at once and wipes what it keeps (you type its name to confirm); its cameras then move to other servers, so use Empty first and wait if you want them moved before the cut. Show removed lists removed helpers.
• Change a helper. Name, Cameras (what it may take), Run the AI on (Auto, GPU or CPU), AI engine (Automatic) and Space for waiting pictures (5 GB, 1 to 100 GB). For a helper somewhere else, Transfer reason shows the legal reason footage may go to it; with none, footage cannot go there. Faces, attendance and number plates stay on the main server.
• Failover. Move cameras off a silent server after (60 s, 20 to 600) moves a silent helper's cameras to a server with room, and the next setting does it sooner (20 s, 10 to 120) for cameras with fire, smoke or person-down checks. Nothing is counted twice. Accept helpers from the open internet (Advanced, off) lets helpers connect from public addresses, and you forward port 7443 yourself.
• Automatic moves. Move cameras automatically (on) lets cameras move to a helper with room when this server stays too busy, and come back later. Move cameras when busier than (80 %, 50 to 95 %, at or under Slow extra checks above) and Bring cameras back below (50 %, 20 to 80 %, at least 10 points lower) set the levels, Busy or quiet before a camera moves (2 min, 1 to 30) sets how long they must hold, and Most automatic moves an hour (6, 1 to 30) stops cameras moving back and forth. Upload this server can give helpers (Advanced, 20 Mbit/s) caps the video a helper may take through this server. Move cameras… on the Helper servers page pins several cameras to a server in one go, shows how much video each sends, and Let the system decide sets them back to automatic. Each helper card also has How it gets camera video, Most busy, Most cameras and Internet speed toward it.
• Where a camera runs. The Cameras list shows "On" and the helper's name, or "Waiting for room" in amber, for a camera that does not run on the main server; recording is unaffected. An event says Checked on and the server.

Storage and retention today

• The Storage page is in the Admin group of the menu and open to every role. It shows free space, what uses the disk, a usage chart with a forecast of when the disk fills, the alerts and the recent activity. Free space is never more than 10 seconds old. The size of each kind of file is measured every 5 minutes.
• The page measures Exports as its own kind of file. It has no Free up space… button, because finished exports delete themselves after 14 days. Edit rules opens Settings → Storage, and the recent activity lists "Old records removed" with what the job removed. Accuracy check evidence has its own line too: check-set pictures are never offered for clean-up, and weekly check pictures are removed after 90 days.
• Rescan (administrators and supervisors) measures every kind of file again within a minute. It deletes nothing.
• Sizes are decimal: 1 GB is 1,000,000,000 bytes.
• Under Recordings by camera, the Keep recordings for slider sets how many days recordings are kept. Cameras that hold the current value follow it; a camera with another value changes only if you tick Apply to all cameras. The page shows the disk the new value needs and what a clean-up would remove. Save changes it, and Undo in the message puts the old values back.
• Alerts appear on the Storage page when the forecast puts the disk at the clean-up line within storage.alerts.warnDays (14 days by default), when recording capacity stays under 90 % of the target for a day, when events waiting to reach the server pile up, and when a measurement fails. The forecast says "Still learning" for the first 24 hours.
• Every 60 seconds the vision worker deletes recordings older than each camera's Retention, 14 days by default and at most 30 under the Strictest privacy profile. While free space is under minFreeGb (50 GB by default, the clean-up line on the Storage page), it then deletes the oldest recordings across all cameras. It never cuts a camera's newest 14 days, and never deletes a segment newer than 10 minutes or one a queued evidence clip still needs. Only the recordings folder is touched.
• Below hardMinFreeGb (10 GB by default) recording pauses on every camera that is not marked Critical, and resumes when free space recovers. A Critical camera keeps recording. Below that floor the Storage page shows a red note, and uploads are refused with HTTP 507, and pictures and clips may fail to save until space is freed. An event whose snapshot could not be saved says Snapshot not saved (disk full).
• Automatic deletion follows only the age rules in Settings → Storage. Legal holds, open incidents, approved samples, datasets and active models are never deleted automatically. You delete an event, a camera (with its events and its review samples) or a model by hand.
• Clean-up shows a preview first: the number of items and the space they would free. Nothing is deleted until you confirm. A large clean-up needs you to type DELETE, and a running clean-up can be cancelled.
• Space saving is set to Auto, On or Off. Auto turns it on when the disk is forecast to fill and off again when there is room. A banner and a chip in the top bar show when it is on. Critical cameras are exempt.
• Each kind of data has its own age in Settings → Storage, and a Never switch keeps it where allowed. No event record goes before the minimum event retention. The site's privacy profile caps each age: under the Strictest profile, event clips and snapshots go after at most 90 days, evidence and event records after at most 3 years, and sightings and visits after at most 90 days. A save above a cap says what the cap is.
• Test bench image results are cleared after 24 hours.

The data folder is laid out like this, which matters when you move or back up the disk:

recordings/<camera>/YYYYMMDD_HHMMSSZ.mp4   60-second segments, named by UTC time (Z marks UTC)
events/<id>/raw.jpg, annotated.jpg, clip.mp4
datasets/inbox, versions/<id>, golden
models/<id>/        training/<job>/        cache/        spool*/

Backups and upgrades

• Backups. A backup holds the database and the files you cannot rebuild: models, branding, the golden set, dataset manifests, and the face pictures and templates. Recordings and event evidence are optional (--with-media). The nightly backup keeps the newest backup of each of the last 7 days and 4 weeks. Copy backups to another machine, because a backup on the same disk is not a backup. A backup never holds APP_SECRET or the export signing key, so keep the secret with your records; after a restore onto a new server, choose Make a new site key in Settings → Sharing and give recipients the new fingerprint. A setup backup (Setup backup) is not a server backup: it holds only the set-up, in a file with its own password, and leaves out events, recordings and the licence. Treat a backup as personal data, because it holds face pictures. Restore refuses an archive that holds a symbolic link. docs/BACKUP.md covers running, scheduling and restoring.
• Upgrades. Plan a maintenance window, because vision stops watching while the images rebuild. Back up, update the code, start, check, and know how to go back. Database changes only move forward, so a rollback restores the backup you took first. docs/UPGRADE.md has the commands.
• Updating an installed package. An administrator opens Settings → Updates and chooses Choose update file (.zip). The page checks the file, then shows what is new and how long the platform will stop. Click Install and enter your password. The page shows four steps (checked, installed beside the current version, switched, checking health) and reconnects by itself after the restart. If the new version fails its checks, the platform goes back by itself, unless that release is marked not to. Then the new version stays installed, the page says so, and you choose Roll back yourself in Settings → Updates or run platformctl rollback on the server. For 7 days Settings → Updates shows Roll back beside the version that is kept. When the update changed the database, rolling back returns it to the backup made before the update, keeps a copy of the newer database for 7 days, and the page says what that loses. On a Windows or macOS install with the built-in database, an update that changes the database layout makes that backup first, and rolling back puts it back. Otherwise it keeps all your data. An update file for another system, an older version or the version already running is refused, and nothing changes. A version older than any installed before is refused even after a rollback. Discard throws a checked file away. If the server refuses its update state, run platformctl repair to rebuild it from what is installed. A packaged server accepts only files signed by the vendor; the first signed release is not published yet.
• Installing a release on a Linux server. An operator installs a signed package on a Docker server with sudo platformctl install --from <file>. The command shows the first administrator password once; if you run it again after a failed first try, it says where that password is kept. A server that runs from the source code moves onto a release with sudo platformctl install --adopt <folder> --from <file>: it asks first, takes a backup, keeps the data, and leaves the old copy in place. The System page calls the store behind live updates Key-value store.

Licence

The Licence page shows what this installation is licensed for. Everyone who can open it sees Status: the state (No licence, Not valid, Trial, Active, Expires soon, Expired with a grace period, or Expired), the customer and site, when it was issued and when it expires, how the AI engine reads it, the cameras used against the limit, and the modules, each On or Locked. Administrators who can manage the licence also see Installation ID, Activate online, Apply a licence file, Online licence and History, which names the source of each change: File, Online activation or Online renewal. The page reads the licence again every 15 seconds.

There are two ways to activate. A server with internet can use a product key. A server without internet uses a licence file. Both give the same licence.

• Product key. Where your system provider has switched it on, the Activate online panel shows. Type the key in Product key (XXXXX-XXXXX-XXXXX-XXXXX-XXXXX) and choose Activate. The box adds the dashes as you type and reads the letters O, I and L as 0, 1 and 1. The server sends only the key, its installation ID and the software version. If the key is wrong, the page says what to check.
• Licence file. The Apply a licence file panel is for servers without internet: upload the .lic file from your system provider, or paste its text. Give the Installation ID to your system provider so the file fits this server.
• Renewal. An online licence renews itself once a day. The Online licence panel shows Last renewed and when it runs until. Choose Renew now to ask at once; the answer is "Licence renewed", "Licence is up to date", or "Online renewal is not set up on this server, so nothing was checked." If a renewal fails, the licence keeps working until it expires, and the panel says "The last renewal failed on" a date and why. The server tries again every day.
• Moving a server. Choose Move to another server…, type the product key and choose Release this server. The server goes back to the basic product at once: up to 4 cameras keep running, the others pause, and no recordings, events or settings are deleted. Then activate the same key on the new server.

A licence sets how many cameras may run, which modules are on (Own branding is one of them), and the product name, logo and accent colour. A trial licence runs for up to 90 days. A paid licence that expires gives a 14-day grace period before paid features lock; a trial key that was released ends at once, with no grace.

If the page says This server has already used a newer licence, the database was probably restored from an older backup. Apply the newest licence again. An online licence also comes back at its next renewal. A licence dated more than a day ahead of the server's clock reads This licence is not valid yet.

• Cameras over the limit. Applying a licence that allows fewer cameras says "Licence applied. It allows 4 cameras: the newest 2 are paused." Those cameras are paused, not deleted: their settings, zones and rules are kept, and the oldest cameras keep running (chapter 6). Status says "2 paused by the licence limit".
• Modules. A feature outside the licence stays in the menu and the command palette, greyed with a lock. Opening it says "<Feature> is not included in your licence." and "Talk to your system provider to unlock." Administrators also see the Installation ID and an Open the Licence page button. Rules and the Test bench need the PPE module, People needs tracking, Datasets and Training need training, and Video search needs search. A refusal by the licence shows one message however many requests were refused.
• Empty pages. An empty page does the same: its next step is greyed out with the reason when your role cannot take it.
• Without the PPE module. Cameras still show live video and keep recording, with motion detection and the disk guard running, but no checks run and no boxes are drawn.
• The AI engine. The engine works out the licence itself. The Status panel reads Agrees or "The AI engine disagrees:" with the reason, and an unreadable or old report reads Not reporting yet. The System page shows a difference in red.
• Brand. The product name, logo and accent colour come from the licence. Settings shows them read-only, and an attempt to change them is refused, unless the licence includes Own branding: then an administrator sets them in Settings, and your values win over the licence's. If the licence lapses, your values are kept but not shown until it is renewed. Reports use the brand in force. The site name is always yours to edit.

A banner at the top of every page tells the state. It says "Trial — 12 days left" on a trial, and for everyone during the grace period "Licence expired — paid features lock on" a date. Administrators also see "Licence expires on" a date, "Licence expired — paid features are off", "The licence is not valid:" with the reason, "Server clock is behind", "The AI engine disagrees:" with the reason, and, for an online licence, "The licence could not be renewed online:" with the reason. The Overview shows a lock, or a retry, if the licence cannot be read.

Privacy desk

Open Privacy in the sidebar (administrators, or anyone given the privacy permission) to answer requests from people about their data and keep the proof. It has seven tabs.

• Checklist. The first tab. It lists what this site has done for privacy and what is left. Each row can carry a Done on date, a Review on date and a note. Nothing here locks anything.
• Requests. Choose New request the day someone asks to see, correct or erase their data. The deadline counts from the received date, and reminders come 3 days before and on the day. Open a request to check who is asking, find their data, export it, erase it and close it. Erasing cannot be undone and asks for your password again. A closed or refused request stays on record for 3 years.
• Consent. Shows each named person and whether consent is recorded or still waited for.
• Holds. Put a person, or a stretch of footage of at most 30 minutes, on legal hold. Nothing about them is deleted or erased until the hold is released.
• Breaches. Log lost, leaked or wrongly seen personal data. The deadline for telling the authority starts when you found it, and reminders come before it passes.
• Documents. Camera signs, worker and visitor notices, the retention policy and other drafts for this site. They are template words: edit them to fit your site before you use them.
• Transfers. Lists helper servers in other countries. Record the reason footage may go to each; a helper server in a country with no reason takes no footage work.

Health check

GET /api/health/deep reports the database, Redis, go2rtc, free disk space, the age of the vision heartbeat, the trainer and the build version. Call it with an administrator session, or with the x-service-token header from the LAN, and point your monitoring at it.

curl -sS -H "x-service-token: $SERVICE_TOKEN" http://localhost:4000/api/health/deep

The status is ok once vision has loaded its model and sent a heartbeat, up to a minute after start. degraded means an optional part is failing: go2rtc, the trainer, low disk or checks.secrets.unreadable. HTTP 503 means the database, Redis, the disk floor or vision is failing.

Remote access

Caddy provides HTTPS on the LAN. For people outside the LAN, docs/DEPLOYMENT.md describes hosting the same web app on Netlify and reaching the on-site API through a Cloudflare Tunnel, with a security checklist to complete first. Remote viewers get live video through the server, about a second behind, because direct video needs port 8555. Their tiles fall back to pictures when that fails.

Troubleshooting

Start with the System page. It shows whether vision, the trainer and each camera are healthy. Items marked IT need access to the server. Logs live in the API's output, recordings/<camera>/ffmpeg.log and training/<job>/train.log.

Getting support

Work through this chapter and the operator guide first. When the problem remains, ask your system provider's support team for help.

1. Note the facts. The time of the problem, the camera or page involved, and what the screen showed.
2. Download support file. An administrator chooses Download support file at the top of the System page. It makes one zip with versions, health, settings, recent warnings and errors, camera and alert-channel status, licence, updates, backups and database row counts. Passwords, keys, email addresses, faces, pictures and people's names are taken out; IP addresses stay, because faults cannot be found without them. Nothing is sent: you choose who gets the file.
3. Send the file with the facts from step 1 to your system provider's support mailbox.

Your system provider sets the priority of each request with these words:

• Priority 1. Alarms are not raised or not sent, or the whole platform is down.
• Priority 2. One or more cameras, or one check, does not work, and the site has a workaround.
• Priority 3. A fault with an easy workaround, or a wrong figure in a report.
• Priority 4. A question or a change request.

Response times are in your support agreement. Never send passwords, keys or tokens, a backup file or pictures of faces. A setup backup also holds password hashes, so keep it as carefully as a password.

Glossary

Absence class

A model class that stands for a missing item, such as NO-Hardhat. It counts as evidence of absence and never as presence.

Accuracy check

A blind weekly set of moments that reviewers answer, to measure how often detection is right on each camera. The answers never change an event.

Alarm

An event from a check that must be acted on at once, such as fire or person down. It shows a banner and a siren on every page until someone acknowledges it.

Area

A named group of cameras, such as a warehouse or a production hall. A camera with no area sits in "No area".

Boxes

The platform's boxes, labels and zones, drawn as a layer over the camera's own video. An event's snapshot comes both ways: Annotated and Raw.

Candidate model

A model that is trained, downloaded or imported but not active. It becomes active only when someone activates it.

Check

One kind of thing the platform looks for in a zone: PPE, a safety check such as fire or person down, or a security check such as a restricted area, an unknown face or a camera blocked or moved.

Checks per second

How many pictures a second the model checks on a camera.

Class map

The table on a PPE model that says what each of its classes means for the platform.

Counting line

A line drawn on a camera's picture. Everyone who crosses it is counted in or out.

Episode

One stretch of time in which one person is missing PPE. It opens one event.

Event

The record of one episode: snapshots, a clip, the missing items, a severity and a review state.

Fine-tune

Train an existing model further on new frames instead of starting from nothing.

Fixed item

Something that looks like fire, smoke or a forklift but never changes, such as a lamp or a steam vent. Checks ignore a still copy of it.

Golden set

Reviewed samples kept as permanent test data. Models are scored on them and never trained on them.

Golden-set score

How well a model finds the items on the golden set, from 0 to 100%. Higher is better; a new site needs 85% to go live without a written reason.

Grace

The seconds a person must be inside a zone before the platform judges them.

Incident

A violation worked as a case, with a status, an owner, notes and a history.

Kiosk

The Monitor wall at full screen with the menus hidden, for a control-room display.

Legal hold

A mark on an event that keeps its evidence. Automatic deletion never removes a held event.

Main stream and sub stream

The main stream is a camera's high-quality stream, and the platform records it. The sub stream is its low-resolution stream, and the platform decodes it for detection, unless the camera is set to read its main stream for detection.

Mining

Queueing frames that look worth reviewing: uncertain, low-confidence or new-looking ones. Its settings are under Frames to review.

Person id

P- followed by eight characters. One person as the cameras re-identify them.

Rule

One PPE type required in one zone, with a severity and an optional schedule.

Sample

A frame in the review queue, with boxes you can correct.

Security mode

A camera with its PPE chip off: it tracks people and judges no PPE.

Severity

High, Medium or Low, set per rule. An event takes the highest severity among its missing items.

Track

One person followed through consecutive frames on one camera.

Version

An unchangeable snapshot of the approved samples, used for one or more training jobs.

Violation

A person missing a required item for long enough to be confirmed.

Visit

One stay of a person in a zone. A return after a long time out of sight starts a new visit.

Warning

A person missing a required item for longer than the rule's warn-after time, but not yet long enough to be a violation. Shown in amber.

Whole frame

The zone every camera gets automatically, covering the entire picture.

Zone

An area of a camera's picture where rules apply.

How the finished system will look

When it is fully built, the platform runs every camera app on one server: recording, PPE, security, safety checks, counting, number plates, video search and attendance, each switched on and off on an Apps page. Below are previews of every screen, drawn from the approved designs and grouped by milestone. Each carries its status: Available, In progress, Next, Planned or Later, and a milestone says Partly available when only some of its screens ship. Screens that are not Available show the plan, and details can change before release.

Today's pages Available

M1: Accounts Available

M2: Client go-live Partly available

M3: Sell-ready white label Available

M4: Security system Partly available

M5: More apps Partly available

Designed screens Partly available

The small details Partly available