Research And System Implementation Of Machine Vision-based Bus Vehicle Section Passenger Flow Statistics Method

In urban traffic management,real-time statistics on the number of boarding and alighting passengers of each line bus at each stop can not only improve the efficiency of the bus department in monitoring and scheduling major bus lines,but also help passengers to get the crowded situation of the bus line they are riding in real time through the Internet or bus stop display,and improve the travel of the public,so it is important to study the method of bus vehicle section passenger flow statistics and realize the passenger flow Therefore,it is of great practical importance to study the method of bus cross-sectional passenger flow statistics and implement a passenger flow statistics system.The main findings of this paper include the following.1)Building a deep learning-based passenger head-and-shoulder detection model.In order to locate the passenger’s position in the video image,a deep learning-based method is used for passenger target detection,and a lightweight YOLOv4-tiny network is used considering the practical deployment of the model.Since the bus scene camera obtains the top-angle passenger video images,the head-and-shoulder labeling method is chosen to build the bus scene passenger detection dataset for training the YOLOv4-tiny network,and experiments show that the model can achieve 99.56% detection accuracy in the bus passenger head-and-shoulder detection test set,and achieve detection accuracy comparable to YOLOv4 in the same task,but the inference speed is substantially improved relative to YOLOv4.2)A multi-target tracking algorithm based on fusion association operator is proposed.In order to solve the tracking instability problem caused by diverse passenger motion behaviors and mutual occlusion of passengers in buses,the fusion association operator is designed to improve the SORT multi-object tracking algorithm for passenger target tracking based on YOLOv4-tiny target detection results using target HSV color space information and location information.The introduced fusion association operator can improve the correlation degree of target information and achieve continuous and robust tracking of passenger targets in complex transit scenarios.3)The bus passenger counting algorithm is proposed.Based on the target detection and tracking results to analyze the motion trajectory of passengers getting on and off the bus in the actual bus scenario,the cross-line counting algorithm and the trajectory classification counting algorithm are designed,and experiments are conducted using the actual bus surveillance video to compare the counting accuracy of the two counting algorithms and verify the counting performance of the studied passenger flow counting system.4)A smart camera-based bus passenger flow statistics platform is established.In order to verify the feasibility of the passenger flow statistical algorithm designed in this paper,a passenger flow statistical terminal is built using a Hessian AI intelligent media processing chip,a web camera and other devices,and the YOLOv4-tiny passenger detection model and tracking and counting algorithm are transplanted into the AI chip to realize the passenger flow statistical function,and experiments are designed to verify that the built passenger flow statistical platform,while ensuring the actual real-time use,has a high The design experiment verifies that the built passenger flow statistics platform has a high statistical accuracy while ensuring the actual real-time use.The paper concludes with a summary of the designed bus passenger flow statistics system and an outlook on the next research.