Safety Monitoring Scheme for Indoor Personnel in Substations based on Helmet Detection and Action Recognition

This paper presents an indoor positioning and action recognition system based on stereo vision, aiming to achieve a safety monitoring system based on helmet detection and action recognition. The system employs the YOLO v8 model for object detection and pose estimation, enhancing the detection capability for small targets by introducing the P2 layer and BiFPN structure, and realizing model lightweighting with the GhostConv module. The BoT-SORT algorithm is utilized for object tracking to ensure the accuracy of action recognition. Action recognition is performed using the ST-GCN, which extracts human skeleton features through a spatiotemporal graph convolutional network. Additionally, a double-threshold detection framework is designed to effectively segment video clips containing the desired actions. Experimental results indicate that the improved YOLOv8 model has enhanced detection performance. This research provides an effective solution for indoor monitoring scenarios, with high practical value and broad application prospects.