A safety behavior monitoring method based on multi-view pose and gaze estimation

This paper presents an integrated framework for safety monitoring by fusing multi-view 3D human pose estimation and gaze analysis. To address occlusion challenges in crowded scenes, we employ a three-stage pose estimation algorithm incorporating associative embedding for robust person disambiguation. For gaze estimation, we utilize head pose as an efficient proxy, enhanced through Euler angle optimization to improve accuracy under large-angle views. The estimated 3D poses and gaze vectors feed into a behavior analysis module that combines rule-based checking and temporal convolutional networks to identify dangerous behaviors like falling, zone intrusion, and inattentiveness. Experimental results on CMU Panoptic and our Industrial Safety Monitoring Dataset demonstrate our method's effectiveness in accurate state estimation and reliable safety alerting, providing a practical solution for intelligent surveillance systems.