The Challenge: 450 Workers, 28 Zones, Manual Spot Checks
A speciality chemical manufacturer operating a COMAH (Control of Major Accident Hazards) regulated site in Teesside employs 450 workers across production, storage, and loading areas. PPE requirements vary by zone — hard hats and safety glasses throughout, chemical-resistant gloves and goggles in processing areas, high-visibility vests and respiratory protection in specific zones. The current compliance monitoring relies on supervisor spot checks conducted four times per shift, covering roughly 15% of the workforce at any given time. Last year, two recordable incidents were linked to PPE non-compliance, and the HSE inspection report flagged inconsistent enforcement as a concern. Each RIDDOR-reportable incident costs the company an average of £45,000 in investigation, remediation, and insurance premium increases.
The site has 28 CCTV cameras already installed. Sending video feeds to external cloud services for analysis is prohibited under the site’s security policy — chemical process data and worker footage cannot leave the site perimeter or UK jurisdiction under any circumstances.
AI Solution: Real-Time PPE Detection from CCTV
A vision model trained on PPE detection processes live feeds from the existing 28 cameras. The model identifies individuals in each frame, detects the presence or absence of required PPE items (hard hat, safety glasses, goggles, high-vis vest, gloves), and cross-references against zone-specific requirements. Non-compliance triggers an immediate alert to the zone supervisor’s mobile device and logs the event with timestamp and camera location for compliance reporting.
Running the detection pipeline on a dedicated GPU server located on-site (or in a nearby UK data centre connected via private link) keeps all video processing within the site’s security boundary. No worker footage leaves the controlled perimeter.
GPU Requirements
Processing 28 camera feeds at 10 FPS for person detection plus PPE classification generates 280 frames per second of inference demand. The model must handle varying lighting conditions, partial occlusions, and workers at different distances from cameras.
| GPU Model | VRAM | Max Camera Feeds (10 FPS) | Alert Latency |
|---|---|---|---|
| NVIDIA RTX 5090 | 24 GB | ~35 feeds | ~1.5 seconds |
| NVIDIA RTX 6000 Pro | 48 GB | ~30 feeds | ~1.8 seconds |
| NVIDIA RTX 6000 Pro | 48 GB | ~40 feeds | ~1.2 seconds |
| NVIDIA RTX 6000 Pro 96 GB | 80 GB | ~55 feeds | ~0.9 seconds |
A single RTX 5090 handles all 28 cameras with headroom for additional feeds as the site expands. Private AI hosting ensures full compliance with the site’s data security policy and GDPR requirements for worker monitoring.
Recommended Stack
- YOLOv8 or RT-DETR for person detection and PPE item classification.
- NVIDIA DeepStream for multi-camera video ingestion and GPU-accelerated decoding.
- TensorRT for optimised inference, maximising frames per second per watt.
- Zone configuration layer mapping camera FOVs to PPE requirement profiles.
- MQTT or Firebase Cloud Messaging for real-time supervisor alerts.
For compliance documentation, add an LLM via vLLM to generate shift compliance reports automatically. Integrate document AI to process worker certification records and PPE training documentation.
Cost Analysis
Manual compliance monitoring using four supervisors conducting spot checks costs approximately £160,000 annually in allocated time. AI monitoring does not replace supervisors but makes their enforcement continuous rather than sampled, increasing compliance coverage from 15% to effectively 100% of monitored zones. The prevented incidents — based on the site’s historical average of two PPE-related incidents per year at £45,000 each — save £90,000 annually. Reduced insurance premiums from demonstrated continuous monitoring typically add another £25,000 in annual savings.
The HSE compliance benefit is equally important. Demonstrating automated, continuous PPE monitoring significantly strengthens the site’s position during regulatory inspections and reduces the risk of enforcement notices.
Getting Started
Capture 5,000 frames from your existing CCTV system across different times of day, weather conditions, and zones. Annotate workers with PPE item labels (hard hat present/absent, glasses present/absent, etc.). Train the detection model and deploy in monitoring-only mode for 30 days, measuring detection accuracy against supervisor observations before activating real-time alerts.
GigaGPU provides UK-based dedicated GPU servers for safety and compliance workloads. Deploy with guaranteed UK data residency and add an AI chatbot for safety procedure queries from the workforce.
GigaGPU offers dedicated GPU servers in UK data centres with full GDPR compliance. Deploy safety detection models on private infrastructure today.
View Dedicated GPU Plans
