Ancilor Vision
Camera events need operational context
A truck on the weighbridge camera is a fact about a place and a time. It becomes useful once it is matched to the delivery it belongs to, checked against the records and reviewed by the person responsible.
A camera sees a truck roll onto the weighbridge at 06:42 and stop. On its own, the event says only where and when. The operational questions are different: which delivery is this, what was expected, what was weighed and received, and who resolves it if those do not agree? Answering them means connecting the event to business records, and being clear about what the event cannot show.
Ancilor Vision is designed to connect camera events with the deliveries and assets they concern, and to bring each one to the person responsible for reviewing it. This article follows one constructed example: a camera event matched to a delivery. The site, times and records are illustrative. The films on the Vision page are licensed stock footage; their tracking boxes and record labels are illustrations made for the site.
A truck at 06:42
Two deliveries are expected at Site D that morning: a supplier load due between 06:00 and 08:00, and a transfer from the depot due between 06:30 and 07:30. At 06:42 the camera over the weighbridge records a vehicle entering the zone and standing for four minutes. The weighbridge prints a ticket at 06:44 with a gross weight, and the gatekeeper enters the delivery document number from the driver's paperwork.
| Record | Comes from | What it can establish |
|---|---|---|
| Camera event | The weighbridge camera | A vehicle entered the zone at 06:42 and stood for four minutes |
| Weighbridge ticket | The scale | A gross weight at 06:44 |
| Document number | The gatekeeper, from the paperwork | Which load the driver says this is |
| Expected deliveries | Ancilor Logistics | What was due at this site, and when |
| Receipt | The receiving site | What was unloaded and accepted |
The match is built from these pieces. The site and the time make both deliveries candidates. The document number points to the supplier load, and the ticket agrees with the camera on the time. Vision proposes the supplier load as the match, the gatekeeper confirms it, and the event is then attached to that load as supporting material. The transfer stays expected.
The camera event is not the delivery
Three things need to stay apart: what the model detected, what the event rule concluded and what the business record says. A detection is a model output on frames. "Vehicle entered the weighbridge zone and stood" is an interpretation of detections over time. "Delivery received" is a business conclusion, and it belongs to the receiving record and the person who signs it.
The camera cannot establish quantity, product or acceptance. The weighbridge gives weight, the receipt gives the accepted quantity, and the site manager accepts it. The event's contribution is narrower and still real: it fixes the arrival time independently of anyone's typing, and it shows when a truck stood on the scale but no ticket was printed.
In this example the event concerns a vehicle, a zone and a time. Nothing in the workflow needs to know who is driving, so nothing in it tries to find out.
From camera to event
Before a camera can add anything to a delivery record, four engineering questions need answers. None of them is about the model.
Getting the stream
Site cameras usually offer their video over RTSP, and many describe their streams, settings and events through ONVIF profiles: Profile S and Profile T for streaming, Profile M for analytics metadata and events. A profile says what a device claims to support, not how a particular camera behaves on a particular network, so each stream should be tested before anything relies on it. Access should be read-only, through an account created for the purpose, and the recorder or video management system the site already runs stays in charge of recording.
Agreeing on the time
Matching an event to a weighbridge ticket works only if the clocks agree. Cameras, recorders, the weighbridge and the server should take their time from the same source over the Network Time Protocol, and the matching rule should still allow a stated tolerance. A camera that runs three minutes ahead of the weighbridge produces events that look plausible and attach to the wrong truck.
Where the processing runs
Video can be analysed on the camera, on a computer at the site or centrally. Processing at the site keeps continuous video on the premises and sends on only events and short clips, but it needs equipment that someone looks after. Central processing is easier to update, and it needs the bandwidth and an agreement on video leaving the site. The choice follows the site's network, its privacy rules and who will maintain the equipment.
Night, weather and obstruction
At night many cameras switch to infrared and lose colour. Headlights glare, rain and fog lower contrast, a dirty lens softens every edge, and anything parked in front of the camera hides part of the zone. An event rule tuned on clear afternoons can miss trucks at night and invent them in heavy rain, which is why it has to be judged on the site's worst conditions as well as its best.
When the match is uncertain
Change one detail: the gate is busy and no document number is entered. Now both deliveries fit the window, and the weight is plausible for either. The event should not be assigned to one of them by default. It stays attached to both as candidates, and the question goes to the gatekeeper with the clip, the ticket and the two expected deliveries side by side.
The gatekeeper confirms which it was, and the confirmation is logged with who made it and when. If the answer later proves wrong, the original event stays and the correction is recorded beside it.
Repeated events need the same care. A truck that stops on the edge of the zone can trigger several entries. Grouping them reduces noise, but a window that is too wide merges two visits into one. The reviewer should see which events were grouped and be able to separate them.
When the camera goes quiet
No event is not the same as no arrival. A camera can be offline, blinded by low sun or blocked by a parked trailer. The review screen should show a gap with its start and end, so that a missing event inside it reads as unknown rather than as absence.
Cameras also move. A camera knocked during maintenance changes what its configured zone covers, and its events can change meaning without any visible error. Moving or replacing a camera should trigger a check of its zones before its events are relied on again.
Who decides, and who sees the footage
The gatekeeper confirms the match. The site manager records and accepts the receipt. The dispatcher follows up an expected delivery that has no event, no ticket and no receipt. A camera event can open a review or, once its match is confirmed, attach to a delivery as supporting material; it does not change stock, close the delivery or release a payment. The responsible person can confirm the event, dismiss it or ask for more information.
Access follows the task. A gatekeeper reviewing an arrival needs the clip of that arrival, not every camera on the site. The purpose of each camera view, permitted users, masking, retention and where video is processed are agreed for each installation. In the EU, the European Data Protection Board's Guidelines 3/2019 on processing of personal data through video devices are the place to start.
Judge it against human review
Evaluate on recorded days from the actual cameras, in the light, weather and traffic of the site, and compare the events with what a person reviewing the same footage records. Count missed arrivals, false events, wrong matches and the time a reviewer spends per case, and agree acceptable limits for this decision before relying on it. Keep the days used for tuning separate from the days used for judging.
NIST's AI Risk Management Framework is a useful companion: a voluntary framework intended to help organisations incorporate trustworthiness considerations into the design, development, use and evaluation of AI systems. The same discipline applies to language models that read operational records; the article on AI-assisted decisions sets out how to test them on cases with a known outcome.
From read-only to review tasks
Begin with one camera, one gate and one question: is this arrival the delivery we expected? Run read-only beside the existing gate process and compare. Only when the queue of open questions stays manageable, and the matches hold up against human review, should events start opening review tasks for the people who own them. The technical reports RATIO and LIBRA set out research behind this approach to review and record-keeping; both propose architectures and report no product results.
References
- Guidelines 3/2019 on processing of personal data through video devicesEuropean Data Protection Board, 2020. www.edpb.europa.eu
- Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1National Institute of Standards and Technology, 2023. www.nist.gov/itl/ai-risk-management-framework
- ONVIF ProfilesONVIF. Cited for the profile names, not as a claim of conformance. www.onvif.org/profiles
- Network Time Protocol Version 4: Protocol and Algorithms Specification (RFC 5905)Internet Engineering Task Force, 2010. www.rfc-editor.org/info/rfc5905
- RATIO: An Architectural Study of Context-Aware Artificial Intelligence in Enterprise SystemsArgjend Bajraktari. Technical report, version 1.0. Zenodo, 2026. zenodo.org/records/23911976
- LIBRA: An Evidence-Centred Architecture for Professional Memory and Decision ContinuityArgjend Bajraktari. Technical report, version 1.0. Zenodo, 2026. zenodo.org/records/23986865
Start with one workflow.
Bring one process, the systems it already runs on and the decision your team needs to make.