Abstract:Aiming at the problem of running algorithms and accurately analyzing pedestrians' poses on mobile low-computing-power devices for intelligent guided tour robots, a pose detection algorithm based on lightweight YOLOv5 is proposed. In the Backbone part of YOLOv5, C3Ghost based on GhostNet is used and in the Bottleneck part, GSConvns and VoV-GSCSP are used to design the thin-neck structure, based on which Slim algorithm and PAGCP algorithm are used for further compression of the model, respectively. Based on the target detection results provided by YOLOv5, the DeepSORT algorithm and SlowFast algorithm are used to realize the tracking of consecutive human targets in the video and the attitude detection of pedestrians. It is experimentally verified that the lightweight and improved YOLOv5 has 71% reduction in FLOPs and 68% reduction in Parameters, and together with the pedestrian pose detection algorithm can accurately recognize human targets and classify pedestrian poses.
