A low-latency AI system for multi-source PPE and fall detection

Abstract:

Workplace injuries in industrial and construction environments remain a persistent concern, often linked to non-compliance with personal protective equipment requirements and delayed detection of fall incidents. Manual monitoring is resource-intensive and may be inconsistent, especially in large-scale operations, which makes automated solutions more practical. A real-time safety monitoring system was developed to detect personal protective equipment usage and human fall events. The system used two separate object detection models based on the You Only Look Once version 8 architecture, one for protective equipment detection and the other for fall detection, both running on a graphics processing unit-accelerated backend. The architecture separated video rendering from model inference. The frontend handled local video display and frame capture, while the backend processed incoming frames and returned structured detection data. Frames were sampled every 200 milliseconds and transmitted through a WebSocket connection. To improve reliability, detections were confirmed only after appearing in multiple consecutive frames, along with a short cooldown period to prevent repeated alerts. The system supported multiple video sources with independent processing and included alert mechanisms through messaging and electronic communication. Detection results were stored for later analysis. The system maintained real-time performance with reduced latency and showed consistent detection behaviour across multiple concurrent video streams.