Human Pose Estimation and Action Classification in Cluttered Industrial Environment for Safety Monitoring

Abstract

This paper presents an automated worker postural classification system using deep learning for real-time activity monitoring at Malaysian industrial sites, particularly fabrication yards and construction sites. The system runs raw CCTV video frames through the YOLOv8x-pose architecture to pull out 17 skeletal keypoints per detected worker. From there, a rule-based geometric classification logic is applied to the lower-body keypoints, specifically the hip, knee and ankle joints, to tell apart standing and sitting postural states without needing the extra computational load that comes with temporal neural networks. The classification logic works through four scale-invariant geometric metrics: a vertical height ratio, a knee flexion angle worked out using the law of cosines, a torso lean and tilt angle and a normalised knee-to-torso proximity measure. All four metrics together build up a solid geometric profile for each action class that stays consistent across different worker body sizes and camera distances. Validation is done using a custom dataset of real-world industrial site footage annotated in YOLO format. The system came out with an overall accuracy of 79.80% and an F1-score of 86.70%. Standing classification came in at 93.00% precision and 90.00% recall, while sitting classification recorded a perfect precision of 100.00% with a recall of 63.20%. The lower sitting recall is mainly put down to lower-body occlusion from on-site equipment and the steep mounting angles of industrial CCTV cameras, both of which block out the lower-body keypoints the geometric logic needs to work with.