AI-Powered Mall Surveillance Integration

The brief

The complex already had a large camera estate and a control room staffed around the clock. The cameras were not the problem - the volume was. A wall of feeds is effectively unwatched, and incidents were being found, but after the fact, rather than as they happened.

The client wanted the existing hardware to do more, to detect incidents in real time rather than only record them for later review, and to do so without adding more staff to the control room.

The challenge

Automated detection is easy to demonstrate and hard to deploy. The difficulty in a live environment is accuracy that operators come to trust.

• False positives are worse than no system at all: an alert stream that cries wolf gets ignored more and more often
• Public and restricted zones need different rules, since normal behaviour in one is an incident in the other
• Inference has to keep pace with the feeds in real time, not batch overnight
• It had to integrate with the existing video management system rather than replace it

How we approached it

We trained detection models on footage from the site itself rather than relying on generic pre-trained weights, because the camera angles, lighting, and crowd density of this particular complex are what the models had to cope with.

Zone rules sit on top of the models. The same detection means different things depending on where it happens and when, so the logic that decides whether to raise an alert is separate from the model that spots the event - and can be tuned by the client without retraining anything.

What we built

Inference runs close to the cameras to keep latency inside the real-time budget, with results routed into the control room's existing workflow so operators did not have to learn a second system.

Alerts carry the frame and the zone that triggered them, so an operator can confirm or dismiss in one look. Foot-traffic patterns are captured alongside the security function, which gave the client operational data they did not have before.

Results

The figures below were measured on the deployed system.

After rollout

Detection thresholds and zone rules were tuned against live operation rather than fixed at handover, since the only honest test of a false-positive rate is a control room actually using it. More on designing around the person who reviews the output in The Future of AI in Enterprise Solutions.