AI Video Analytics Platform

AI Video Analytics Platform

This platform turns ordinary CCTV streams into operational alerts: PPE compliance, safety events, and vehicle analytics. Footage stops being something nobody watches until something goes wrong. I built it as a per-camera system, with fleet health monitoring built in.

Challenge

Industrial sites, construction sites, and factories run cameras for security, not for analysis. Footage sits unreviewed unless an incident forces someone to go looking. Site owners actually needed PPE compliance turned into an operational signal: who's missing a hardhat or vest, where, and when. They also needed person-safety events like falls, without adding another screen for someone to babysit all day.

A system that watches camera streams around the clock also has to watch itself. A camera or a worker process going dark silently is worse than no system at all.

Solution

I built a per-camera worker architecture. Each RTSP or webcam stream runs its own process, feeding a detection and tracking pipeline. That pipeline dispatches events through a supervisor to an audit log and a webhook.

Detection models were chosen for commercial-safe licensing. The object-detection family most teams reach for by default is AGPL. AGPL blocks closed-source commercial deployment unless you buy a vendor license. The stack here runs on permissively licensed alternatives instead.

Camera placement mattered more than any model choice. Helmets and vests read reliably from standard CCTV. Glasses and footwear don't, so those need checkpoint cameras at chest-to-head height. That's where people pass single-file, a few meters out.

False positives are controlled with deduplication: one person passing counts as one event, not one per frame. A per-track cooldown and a review queue catch borderline detections instead of alerting on everything a model is unsure about. Every alert ships with a snapshot and a short clip, so a supervisor sees evidence, not just a claim. I go deeper on the deployment side in PPE detection cameras: what a real deployment requires.

Camera fleet health runs alongside detection. Each camera gets a heartbeat, with an alert when one goes silent past its threshold. Process respawn is crash-safe, so one dead stream never takes the rest of the fleet down with it.

Impact

Supervisors get PPE and safety events as evidenced alerts: photo, timestamp, location. That replaces reviewing hours of footage after the fact. The same pipeline covers person-safety events and vehicle analytics: zone counting, parking overstay, plate reads. One platform now replaces separate tools for security, safety, and site operations.