A Lightweight Method Integrating Keypoint Detection and Perspective Geometry for Substation Safety Distance Monitoring

Abstract

In the high-voltage environment of substations, due to the lack of effective monitoring methods, personnel may fail to maintain a safe distance from the equipment due to subjective factors, leading to electric shock accidents. To more effectively reduce the occurrence of such accidents, this paper proposes a monocular vision-based method for monitoring the safety distance between personnel and live equipment. Firstly, this paper makes lightweight improvements to the YOLOv8 (You Only Look Once) keypoint detection model by introducing the proposed dense feature fusion (DFF) module and adaptive channel cross (ACC) module into the backbone and neck, respectively, and replacing the network's original path aggregation network (PAN) structure with the bi-directional feature pyramid network (BiFPN) structure. Subsequently, based on the principles of camera imaging and perspective geometry, this paper calculates the mapping relationship between two-dimensional pixel coordinates and three-dimensional world coordinates. Finally, based on the detected keypoints and the obtained mapping relationship, the distance between personnel and live equipment is monitored. Experiments conducted in a substation scenario show that the improved keypoint detection model increases AP from 0.718 to 0.771, reduces parameters from 3.09M to 2.10M, and lowers FLOPs from 8.48G to 7.01G, and the maximum distance measurement error is only 3.778%.