Statistal Methods of Bus Passenger Flow Based on Improved YOLOv5s and DeepSORT Algorithms

Aiming at the problem of low tracking accuracy in the statistics of bus passengers getting on and off, an improved passenger flow statistics algorithm based on YOLOv5s combined with DeepSORT is proposed. The GAM integrated into Group Convolution and ChannelShuffle is integrated with the backbone and neck parts of the YOLOv5s network structure to enhance the network feature extraction capability. The Decoupled Head is replaced by the YOLOv5s detection head to improve the detection accuracy of passengers in crowded situations. Alpha-IoU is used as the target frame regression loss to further improve the positioning accuracy. The improved YOLOv5s is connected with DeepSORT, and detection lines are set up in the monitoring video of the front and rear doors of the bus to count the flow of passengers getting on and off. The experimental results show that our algorithm in this paper is 1.1% and 8.5% higher than the YOLOv5s algorithm in mAP0.5 and mAP.5:.95, respectively. After connecting DeepSORT, the passenger flow statistics in the scene of getting on and off the bus, the accuracy rate reached 96.8% and 98.1%.