Video Analytics and Surveillance AI
Turning camera feeds into a small number of alerts an operator will actually act on, with the privacy and lawful-basis work done before the cameras are pointed anywhere.
The reason most video analytics deployments fail is not detection accuracy. It is that they generate more alerts than anyone can look at, so within a month the operators mute them, and the system continues running while nobody watches it.
Video analytics is the automated interpretation of camera feeds to detect events, track objects and people across time and cameras, and raise alerts or produce statistics, deployed with an alert threshold tuned to what operators can genuinely act on and with privacy controls designed in rather than added later.
Alert fatigue is the failure mode
An operator monitoring a wall of cameras can act on a small number of alerts per hour. If the system produces more, the surplus is not extra vigilance; it is noise that trains people to ignore the system entirely.
So we work backwards from capacity: how many alerts per operator per hour are actionable, what precision that implies at your event rate, and which event types are worth that budget. Detections that do not clear the bar become statistics rather than alerts, which is often where their real value was anyway.
Count the alerts before you build
A quick calculation on historical footage usually settles the design. A detector with excellent precision on a rare event still produces an unusable volume if it runs across two hundred cameras continuously, and knowing that in week one changes what gets built rather than what gets blamed.
What we build
Detection tuned to your scenes
General detectors underperform on real camera positions: high angles, wide lenses, poor light, weather and occlusion. We fine-tune on footage from your own cameras and evaluate per camera, because performance varies enormously between them and an average conceals the ones that do not work.
Tracking within and across cameras
Maintaining identity as an object moves through a scene and between cameras is where most of the difficulty sits. It is what enables dwell time, path analysis and follow-through, and it is considerably harder than detection, particularly in crowds.
Events defined in operational language
Not 'person detected' but 'person in the exclusion zone for more than ten seconds during operating hours'. Rules with zones, durations, directions and schedules are what turn detections into something worth interrupting somebody for.
Privacy designed in
Blurring or masking by default, processing at the edge so raw footage does not travel, retention limits applied automatically, and access controls with an audit trail. In many jurisdictions these are legal requirements rather than good practice, and retrofitting them is considerably harder than including them.
Operator interface built for triage
The clip, the reason, the confidence and one-click confirmation or dismissal, with dismissals captured as training data. An alert that requires scrubbing through footage to understand will be dismissed unexamined.
Statistics as the quiet benefit
Counts, dwell, flow, occupancy and queue length are frequently more valuable than the alerts and carry none of the fatigue risk. Many deployments end up justified by the operational analytics rather than by the incidents they catch.
Lawful basis and public trust
This is the area where we most often recommend legal and, where relevant, works council engagement before anything is installed:
- A documented lawful basis for processing, and a data protection impact assessment where the processing is high risk.
- Signage and transparency appropriate to the setting, and a defined process for subject access requests including footage.
- Where employees are monitored, consultation obligations that in several jurisdictions are a precondition rather than a courtesy.
- Special caution around biometric identification, which is separately regulated and covered under facial recognition and biometrics.
- Retention that is enforced automatically, since 'we keep it for thirty days' is a policy until it is a scheduled deletion job.
We build these controls in and hand over the evidence, because the questions arrive before go-live and are considerably cheaper to answer with a design than with a promise.
How the engagement runs
Alert budget and privacy design are settled before detection work begins.
Scene audit and alert budget
Camera positions and footage assessed per camera, actionable alert capacity established, event types prioritised.
Privacy and legal design
Lawful basis, impact assessment, masking, retention and access controls agreed before deployment work.
Detection and tracking
Models fine-tuned on your footage, tracking within and across cameras, evaluated per camera rather than in aggregate.
Event rules and operator interface
Zones, durations and schedules configured, triage interface built with dismissal capture.
Pilot and handover
Limited camera pilot with alert precision measured against operator dismissals, then rollout and handover.
What you receive
Alerts operators act on, statistics they use daily, and the privacy evidence to justify both.
Detection and tracking system
Deployed at edge or centrally, tuned per camera, with measured performance for each.
Event rule engine
Zones, durations, directions and schedules configurable by your team.
Alert precision report
Precision at the operating threshold per event type, with the alert volume per operator hour.
Privacy pack
Lawful basis, impact assessment, masking, retention enforcement and access audit.
Operator interface
Clip, reason, confidence and one-click adjudication with dismissal capture.
Operational analytics
Counts, dwell, flow and occupancy reporting, with the dashboards your operations team will use.
Is this the right engagement?
Worth being direct. Video Analytics and Surveillance AI is the wrong spend in some situations, and those are listed rather than buried.
Good fit if
- Cameras exist and footage can be accessed for evaluation.
- There are specific events worth interrupting an operator for.
- Someone will define what an actionable alert is and how many are tolerable.
- Privacy and legal can engage before deployment.
- Operational statistics would be useful even if alerts were not.
Choose something else if
- There is no lawful basis for the monitoring proposed.
- Camera positions or quality make the events genuinely invisible.
- Nobody is monitoring alerts, so there is no one to alert.
- The requirement is biometric identification, which is facial recognition and biometrics.
Frequently asked questions
Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.
Why do video analytics deployments get switched off?
Alert fatigue, almost always. A system that produces more alerts than operators can act on trains them to ignore it within weeks. We work backwards from actionable alert capacity and turn everything below that bar into statistics rather than interruptions.
Can it track people between cameras?
Yes, and it is considerably harder than detection, especially in crowds and with varied camera angles. We measure tracking quality per camera pair rather than assuming it, because it is what dwell, path and follow-through analysis depend on.
Is video analytics legal where we operate?
Monitoring for safety and security is common and lawful in most jurisdictions with the right basis, signage and retention limits. Biometric identification is separately and more strictly regulated, and employee monitoring often carries consultation obligations. We involve legal before installation rather than after.
Does footage have to leave our site?
No, and often it should not. Processing at the edge means only events and metadata travel, which reduces bandwidth cost and simplifies the privacy position considerably. That deployment pattern is covered under edge computer vision.
What if our cameras are old and low resolution?
Then the honest answer may be that certain events are not detectable, and we establish that on your real footage in the first fortnight. Camera replacement is sometimes the correct recommendation, and it is better made as a finding than discovered after a model has been trained.
Often paired with this
Most clients combine two or three engagements from the Computer Vision pillar. These are the ones that most often run immediately before or after.
Pose Estimation and Activity Recognition
Movement and activity understood over time, using skeletal data that needs no identifiable footage.
Read more →Facial Recognition and Biometrics
Biometric systems built lawfully, tested across demographics, with liveness, template protection and a narrow scope.
Read more →Optical Inspection for Edge Devices
Vision models running on device within real thermal, power and latency budgets, with fleet updates handled.
Read more →Is this the right engagement?
Tell us what you are trying to build. If a different service fits better, or if you do not need us at all, we will say so.