Kamal.Case study
← Selected work
Case study · Computer vision · 2024

Safety compliance detection on a live factory floor

Industrial real-time PPE detection framework

ClientBit Bridge AI
Engagement2 weeks, Dubai Expo
RoleCV model + integration
VerticalManufacturing

A real-time AI safety monitoring system for manufacturing environments. It reads existing industrial CCTV, detects helmet and safety-vest compliance frame by frame, and raises an instant alert the moment someone walks onto the floor without them — no new hardware, no guard watching a wall of monitors.

The problem

Compliance was a clipboard exercise.

PPE rules on a manufacturing floor are only as good as the person checking them. Spot audits catch a fraction of violations, cameras already recording everything are watched by nobody, and by the time a near-miss is reviewed the footage is a week old. The client needed the cameras they already owned to do the checking — continuously, and loud enough to act on.

Solution architecture

Three moving parts, one existing camera network.

01
Video input

Live CCTV streaming straight off the industrial camera infrastructure already installed on site.

Real-time feedsMulti-camera supportIndustrial-grade capture
02
AI processing

A custom computer-vision model integrated with Frigate NVR, inferring on every frame as it arrives.

Bounding-box detectionHelmet & vest recognitionReal-time inference
03
Alert system

Instant notification the moment a violation is seen, with the state visible on screen for the floor supervisor.

Real-time alertsCompliance trackingVisual indicators
Detection showcase

What the model sees.

Helmet detection
Helmet detection

Real-time identification with bounding-box visualisation, tuned for hard hats under mixed factory lighting.

Full PPE compliance
Full PPE compliance

Automated validation of both hard hat and high-visibility vest, resolved per person on a wide-angle floor feed.

Technical stack
Video managementFrigate NVR
AI frameworkCustom CV model
Detection methodBounding box
Input sourceIndustrial CCTV
Processing typeReal-time stream
Industry verticalManufacturing
Key capabilities
Multi-object detection

Simultaneous helmet and vest identification per person in frame.

Live stream processing

Continuous monitoring with minimal latency between event and alert.

Bounding-box visualisation

Clear on-screen indicators so a supervisor can see what the model saw.

Compliance alerts

Instant violation notifications, tracked over time rather than one-off.

Outcome
2 weeks
Concept to deployed
24/7
Monitoring coverage
Real-time
Detection speed
0
New cameras bought
What I'd carry forward

Honest notes

The hard part was never the model — it was the frame budget. Running inference on every camera at once meant choosing what to skip, and being honest about the false-positive rate rather than tuning it away in a demo. A supervisor stops trusting an alert system the second it cries wolf twice.

Similar problem?

I build vision systems that survive the shop floor.

If you have cameras already recording and nobody watching them, that's usually a two-week conversation. Send a short brief and I'll come back with scope, timeline and price.