Method And Application Of Bus Passenger Flow OD Matrix Acquisition Based On Deep Learnin

The public transportation passenger flow OD matrix records the flow data of passengers’ boarding and alighting points,serving as a crucial basis for the planning and management of the public transportation system.By studying the OD matrix,it is possible to ensure the scientific formulation of operational plans and the rational allocation of bus resources,as well as evaluate the operational efficiency of the public transportation system and the actual effects of policies.In response to issues such as insufficient accuracy,poor efficiency,and limited functionality of existing methods,this paper proposes a deep learning-based method for obtaining the bus passenger flow OD matrix.This method considers both speed and accuracy by utilizing color cameras installed on the top of the front and rear doors of buses to capture videos of passengers boarding and alighting.Deep learning techniques are employed to extract the passengers’ positions and motion trajectories from the videos,enabling passenger counting functionality.Furthermore,the method matches alighting passengers with boarding passengers,allowing the acquisition of the passenger OD matrix.The specific details are as follows:(1)Considering the characteristics of deploying passenger detection methods on edge computing devices inside buses in the application scenario of this paper,we propose the YOLO-Mobile object detection method.This method consists of four parts: backbone,neck,head,and post-processing.The backbone utilizes the Mobile Net V3-Large feature extraction network and the SPPF structure to reduce the required computational and storage resources.It also employs a fusion method that combines the input image with the predictions from a Kalman filter to improve detection accuracy.The neck employs the Bi FPN-Mobile structure,the head uses a decoupled architecture and anchor-free design,and the post-processing utilizes the Sim OTA post-processing method to further enhance accuracy.The application of these methods enables YOLO-Mobile to achieve a certain balance between accuracy and speed.(2)Building upon the previous step,we further investigate the problem of multi-object tracking.Considering the limitations of the SORT multi-object tracking method,which relies solely on the motion features of targets and can lead to ID switches,as well as the characteristic of passengers in the application scenario of this paper,where passengers do not disappear or appear within the recognition area,we propose the SORT+ multi-object tracking method.This method employs a cascaded matching approach to determine whether the targets and tracking trajectories are normal.For abnormal targets and tracking trajectories,the method utilizes the appearance features of targets to assist in matching.(3)Building upon the passenger detection and tracking methods,we further investigate the problem of re-identification.Considering the characteristics of deploying reidentification methods on edge computing devices inside buses and the need for frequent matching of passengers,we propose the PRe ID passenger re-identification method.This method consists of three parts: backbone,head,and post-processing.In consideration of the characteristics of edge computing devices,the backbone utilizes the Mobile Netv3-Large feature extraction network and the SPPF structure.The head adopts a decoupled structure to achieve re-identification feature clustering and passenger attribute recognition.The post-processing stage employs a fast matching method based on passenger attributes to improve matching efficiency.