Development of an AI-Based Awkward Working Posture Detection System Using the REBA Method with Edge Computing and Deep Learning Approaches in Industrial Workshops

Awkward working postures pose ergonomic risks in industrial workshops, particularly when combined with force, repetition, or prolonged exposure. This study implements a human-in-the-loop ergonomic assessment system integrating MediaPipe Pose, Rapid Entire Body Assessment (REBA), and edge computing. A Jetson Orin Nano processes uploaded media and Real Time Streaming Protocol (RTSP) camera streams. MediaPipe estimates 33 body landmarks to calculate six segment angles, while HSE personnel verify landmark placement and task factors before final scoring. The system stores assessment records in MySQL and displays them through a web dashboard. Four workshop postures involving low-workpiece handling, pipe maintenance, grinding, and welding demonstrate system functionality. System-derived REBA scores were 5 for squatting (Case A) and conditionally 5 for bending (Case B). Case B's final score was withheld because wrist visibility fell below 0.50. Observer scores were 7 and 5, respectively, with both cases classified as medium risk. Grinding and welding cases were evaluated without observer comparisons. Segment agreement reached 3/5 for Case A and 5/5 for Case B. During a two-hour RTSP test involving 93,729 frames, mean latency was 74.14 ± 19.64 ms, throughput reached 13.04 ± 2.08 FPS, and CPU/GPU utilization averaged 21.35%/57.71%. Findings demonstrate operational feasibility while identifying hip-landmark placement and occlusion as key sensitivity factors. However, these descriptive results neither establish measurement accuracy nor demonstrate reduced musculoskeletal disorder incidence. Further instrument-based validation across diverse postures and operating conditions remains necessary to establish measurement reliability and broader practical applicability.