Research Of Bus Passenger Flow Statistics Technology Based On Video Image Processing

Passenger flow statistics is of great significance for real-time bus scheduling and optimization of bus line operation.Due to the cameras in the passenger area of the bus,the existing system generally adopts the scheme based on video image processing.Specifically,the passengers appearing in the video are detected and tracked separately,and then determining the behavior of getting on and off the bus according to the motion trajectory,thus the statistical count is completed.In recent years,with deep learning constantly in terms of object detection,this thesis proposes to apply deep learning technology to passenger detection in buses.Designing multi-target tracking algorithms and trajectory analysis methods combined with distance information,to solve the low accuracy problem of the existing system and widened the use range of the application.This thesis first introduced related technologies,parameters,and hardware and software platforms.Then design a deep learning algorithm with the passenger head as the detection target,and proposed an acceleration method based on cpu-fpga architecture to improve the inference speed in the embedded environment.Next,a multi-target tracking by detection algorithm is designed,which use IOU to finish data association method.This method reduce the rate of tracking loss under crowded conditions.For motion trajectory analysis,a method of combining the distance and the reference line is provided to adapt to different camera positions.Based on the above algorithm design,a prototype system of flow statistics has completed on the ZYNQ platform.In addition,test the performance of each part of the algorithm.The test results show that the system can achieve average accuracy of 87% for bus traffic statistics,and achieve 80% accuracy in other similar public scenes.