Dynamic Video Streaming Real-time Headcount Based on Improved YOLO V5 for Open Plazas

Abstract

The difficulties of dense and dynamic movement of people can cause shelter and deformation, as well as affected by light. This paper proposes an effective measure for the detection tracking and counting of dynamic pedestrian volume in open plazas based on YOLO V5. The small target detection layer is added with the deep learning network YOLO V5, and the Cross Stage Partial (CSP) module is improved. Attention mechanism and bi-directional feature pyramid network are added, too. The experimental results show that the proposed method has high detection speed and low missed detection rate. It can accurately detect and calculate the total pedestrian flow in the open plazas. The method also has accurate identification of pedestrians in different light and weather conditions, as well as ensures real-time detection.