A Method Study For Bus Passengers OD Recognition Based On Computer Vision

Origin and Destination information of passengers is an important basic data for bus operation management and planning.It is of great significance to analyze and improve the operation efficiency of the public transportation system.The traditional method of manual investigation has the disadvantages of high cost and low sampling.In recent years,smart card data has been widely used to extract OD information.However,smart card data lacks passenger disembarkation information,and there is a certain bias in the use of smart card data,which reflects certain deviations in the true travel behavior of residents.The bus monitoring system has been widely used in buses,and the monitoring video data provides new ideas for obtaining information on real passengers getting on and off the station.Video data is relatively easy to obtain and covers a wide range,making up for sample deviations and unverifiable problems caused by IC data.With the development and maturity of video processing technologies represented by deep learning,it is possible to detect and track targets from video data and achieve cross-camera target recognition.Therefore,this paper studies the method of using computer vision technology to obtain the pair of bus passengers getting on and off the bus.The specific work and innovations of the paper are as follows:(1)Based on the YOLOv3 target detection framework,the target detection of bus passengers is realized.In order to improve the performance of target detection,we use the video captured in the bus,through the frame framing and extraction method to obtain the video shooting image,use the annotation tool to manually mark the passenger object,and obtain the bus passenger data set.Perform cluster analysis on the data set to obtain a pre-selection frame based on human targets,and finally use the bus passenger data set to train to obtain a target detector for bus scenarios.Finally,the m AP value of the detector reaches 89.4,which is 12.3% higher than the original target detector.(2)Track the detected target and extract passengers on and off based on the tracking trajectory.The Kalman filter algorithm is used to predict the state of the detection frame,and the three matching methods of apparent feature,motion information and cascade matching are used to achieve more accurate matching accuracy and achieve continuous tracking of multiple passenger targets;Obtained tracking object,we use the tracking trajectory to determine whether the detection object has the behavior of getting on and off,and use the threshold method to determine whether it belongs to the getting on or off behavior,so as to filter the detection and tracking objects,and get the passenger data set and getting off Passenger data set.(3)Re-identify passengers who get on and off the vehicle to obtain passenger OD information.The re-identification algorithm is based on the PCB+PRR framework,and uses a random erasure algorithm to train the training data set.Compare the effect of random erasure,adding verification set and other methods on the training results.At the same time,based on the bus scene,this paper filters and screens the data set for re-identification,so as to reduce the detection time of re-identification and improve the detection accuracy.Finally,the filtered and screened passenger data sets are re-identified,and finally the bus passengers get on and off site pairs to obtain passenger OD information.Through comparative experiments,we get the best performance of the re-recognition algorithm on public data sets,with Rank-1 reaching 93.78 and m AP reaching 83.62.This algorithm is applied to the bus passengers getting on and off data set,and the matching accuracy rate is 80.4%,indicating that the algorithm has a good application effect for identifying cross-camera bus passengers.