How it works
Video goes through six steps: connect your streams or recordings, detect people, vehicles and bags, describe what can be observed, decide what matters with rules, put each alert in front of a person, and pass the results to the tools your teams already use.
- 1Detected · Car, followed on this camera
- 2Described · White car, plate read
- 3Seen again · A candidate at the next camera
- 4Kept · A still and a record to search
The six steps
From a camera stream to a reviewed alert.
What happens at each step, and what you get from it.
Connect
Read what you already have.
- Live streams from IP cameras, recorders and video management systems that publish a standard stream
- Recorded footage uploaded as single clips or several at once
- One connection per camera, read from inside your network
- The existing recorder keeps recording; VisionObject reads alongside it
Detect
Find people, vehicles and carried bags.
- Frames are sampled from each stream and analysed
- People, cars, motorcycles, bicycles, buses, trucks and bags are detected and followed within one camera
- The clearest frame of a sighting is kept as a still, where one was captured
- Uploaded footage is analysed as fast as the hardware allows, at the time it was recorded
Describe
Record what can be observed.
- Colours and short descriptive tags, where the appearance module is on
- Plate text with a confidence value, where the plate is legible
- An AI scene reading of one still frame, where that module is on
- Sightings that look alike grouped as candidates, on request
Decide what matters
Rules, not motion.
- Zones and lines drawn on each camera, with dwell limits, and count thresholds for a camera's whole view
- Schedules, severity and repeat limits on every rule
- A plate watchlist kept by each organisation
- Rules run on live cameras only; uploaded footage never raises an alert
Review
An alert stays New until a named person marks it.
- Each alert carries its camera, time and rule, and leads to a still where it has one
- A named person marks it Acknowledged or False positive
- Alerts can be assigned and annotated
- Changes, and searches by plate, scene description and route, are written to the audit log
Integrate and report
Use the results in your own tools.
- An authenticated API for sightings, events, rules, alerts and plates
- A real-time event stream your systems can subscribe to
- Counts, the alert funnel and camera health in Insights
- A weekly brief for each organisation, where that module is on
Four paths
Live viewing and analysis are separate, and each reports its own state.
A camera can be showing live video while its analysis is behind, or the other way round. VisionObject keeps the two apart and says which is which.
- The live path
- The picture an operator watches. It does not wait for analysis.
- The analysis path
- Sampled frames, detection, description and rules. When it falls behind, old frames are dropped rather than analysed late.
- The real-time alert path
- Zone, line, dwell and count rules, and the plate watchlist, are decided from the analysed frames without waiting for any AI scene reading.
- The reading path
- An AI scene reading is made after a sighting. Rules that depend on it follow the reading, and stop if the AI service is unreachable.
Where the time on a record comes from
Live sightings carry the time the frame reached the server, not the clock inside the camera. Uploaded footage takes its start time from the file name where the export includes one, otherwise from what the operator enters.
Platform map
Sources in. Reviewed results out.
Streams and recorded footage go in. Investigators, operators, managers and your own systems read a searchable record of what was seen.
Sources
- IP camera streams
- Recorder and VMS streams
- Uploaded footage
- Uploaded images
VisionObject
Video analytics engine
- Detect and follow people, vehicles and bags
- Describe colours, plates and tags
- Collect events of interest
- Apply rules, schedules and severity
Searchable index
Sightings, events, stills
- Type, colour, plate, tag
- Camera, zone, time
- Similar appearance
- Stills of sightings
- Retention you set
Outcomes
Investigate
Search and filters
Alert
For a named person to review
Report
Insights and the weekly brief
Integrate
API and event stream
Frame to alert
What happens to a frame.
Select a step to read what it does.
Step 1
Capture video
Cameras record the scene, or you upload footage.
Live streams from IP cameras, recorders and video management systems, and exported clips you upload. Nothing is installed on the cameras, and your recorder keeps recording as before.
Step 2
Bring it in
The platform reads each feed from inside your network.
Step 3
Prepare frames
Frames are sampled and time-stamped for analysis.
Step 4
Analyse
People, vehicles and bags are detected, followed and described.
Step 5
Show results
Results appear in search, the alert queue and Insights.
Step 6
Review and act
A person assesses the evidence and decides.
Modules
Switch on what you need, for each organisation.
Detection and tracking always run. Appearance, number plates, AI scene reading and the weekly brief are switched per organisation. Face search is not offered in the current release. Optional does not mean experimental: each module's prerequisites and limits are in the availability register.
Modules
Detection and tracking always run. Every other module is switched on or off for each organisation; switching one off stops it running and hides what it produced.
Detection and tracking
Always onPeople, vehicles and carried bags found in each analysed frame and followed within one camera.
Appearance
Optional, per organisationColours, short descriptive tags, and grouping of look-alike sightings across cameras.
Number plates
Optional, per organisationPlate text and a confidence value on vehicles large enough to read, plate search, routes and the plate watchlist.
AI scene reading
Optional, per organisationA short description and up to eight tags for a sighting. Sends one still of the whole camera view per sighting to the AI service you configure, with the object type, the plate as read, the camera's identifier, colours, how long the object was followed and the time; a failed request may go to a second configured service.
Weekly brief
Optional, per organisationA weekly fact sheet with a written summary. Sends aggregate figures, with the names of your rules, cameras, areas and watchlist labels, to the AI service you configure.
When something fails
What stops, and how you are told.
A camera feed drops
The platform retries with a growing delay, shows the camera as degraded and then offline, and raises a camera-offline event. Nothing is analysed from that camera until the picture returns.
Analysis falls behind
Live viewing and analysis are separate paths. When analysis cannot keep up, the oldest waiting frames are dropped and frames that are too old are skipped rather than analysed late. Each camera shows whether it is live, behind, stalled or lost.
The analysis machine runs short of graphics memory
Optional analytics such as plate reading and appearance pause for about a minute and the system reports itself as degraded. Detection continues. An alert that depends on a paused module, such as a plate watchlist hit, is not raised while it is paused.
The AI service is unreachable
Scene readings stop, unless a second AI service you configured can take them. Readings that have waited too long are skipped, not replayed, and rules that depend on them do not fire. Detection, zone and line rules carry on.
A part of the system stops
Each part reports its health, and the deployment shows as unhealthy. VisionObject runs on one machine: if that machine stops, analysis stops until it restarts. Your cameras and recorder are unaffected.
Security and privacy
Inside your environment, under your control.
Runs in your environment
On a workstation or, after assessment, on a server or a virtual machine in your own cloud account. Camera streams are read from inside your network and are not published to the internet.
Everyone signs in
Each person has an account, one of five roles, and an optional list of the cameras they may see. Changes, and searches by plate, scene description and route, are written to an audit log with their name.
Retention, and what has none
Sighting records and stills are deleted after 90 days by default, a period a platform administrator can change. Frame-level detection data is deleted after 30 days, which is fixed. Plate-read history, events, alerts and their notes, weekly briefs and uploaded footage have no retention period in the current release.
Face search is not offered
It is not in the current release: the recognition model has to be replaced with one licensed for production use first.
See it working on footage like yours.
Tell us the workflow you care about. We will walk through it on demonstration footage, say what is available today, and plan a pilot on your own cameras.