A research team has built a single GPU-based framework that runs six surveillance tasks at once — and makes the results reliable enough to act on.
The system handles face recognition with zone-based access control, license plate reading, weapon detection, fire and smoke alerts, and human action recognition simultaneously. Two of its models were built from scratch to fill gaps in public datasets: a weapon detector that hit 94.7% precision (mAP@0.5) and a vandalism classifier trained on 614 video clips that reached 94.33% accuracy. The framework runs on commodity hardware with per-frame latency under 100 ms, meaning it qualifies as real-time by most operational definitions.
The more interesting engineering is what happens after detection. A temporal validation layer — combining multi-frame confirmation, confidence-weighted voting, and cascaded filtering — converts noisy per-frame guesses into confirmed events. That step cut the fire and smoke false-alarm rate from 52% to 4%, and pushed license plate exact-match accuracy from 66.7% to 81.8%. Those are the numbers that determine whether an operator trusts the system enough to act on it.
Surveillance AI research tends to optimize for detection metrics and leave the false-alarm problem as someone else's issue. This work treats reliability as a first-class design requirement — which is either a sign of maturity in the field or a pointed comment on how little the field has demanded it until now.