Real-time people localization for safe-zone supervision

Abstract

A privacy-conscious, on-premise, multi-camera system that supervises an industrial safe zone is created and tested on an indoor test platform using fixed RGB colour cameras and edge computing. The system detects people and personal protective equipment (PPE), and a rule-based PPE compliance decision is made in the camera view covering the entrance gate to the platform. People are localised on a single 2D floor map by estimating a point of ground contact and mapping image measurements to the ground plane through calibrated homographies, followed by multi-view fusion and anonymous tracking to maintain consistent track IDs across the cameras without biometric identification. Intrusions into restricted (no-go) zones are identified on the floor plan and reported as event-level alerts. Quantitative performance on annotated sequences demonstrates high accuracy in person detection and segmentation, a high degree of consistency in safety-critical compliance decisions, multi-camera tracking with stable anonymous identities under typical operating conditions, and localisation errors on the floorplan below the sub-metre range in the vast majority of observations. Detection is nearly perfect at the event level with a low false-positive alarm rate, and end-to-end performance is real-time on edge hardware with multiple high-resolution video streams at low latency. The findings show that an edge-only design can be used to attain credible safety monitoring and zone supervision.