Research And Application Of Vehicle Tracking Algorithms Based On Video Stream

Target tracking is a hotspot and difficulty in the field of computer vision for many years.Although substantial progress has been made in recent years,even the most advanced multi-target tracking algorithms are still challenged by deformation,background interference and mutual occlusion between targets.In urban traffic scenarios,vehicles tracking can provide intelligent operation for traffic monitoring system.In this paper,we use single target tracker combined with data association to achieve multi-target tracking task,and apply it to traffic monitoring scenarios.Finally,cluster vehicle trajectories according to the tracking results.A single target tracker is set for each target in the algorithm to reduce the dependence of system performance on the detector.In the single target tracking algorithm,the extraction of moving vehicle features is very important.In order to improve the robustness and accuracy of the algorithm,the target appearance features extracted by deep learning model.By analysising the vehicle trajectories,it is considered that the vehicle trajectory can be fitted into a curve or even a straight line in a short time.Based on this,the direction and distance of the vehicle motion can be predicted,and can reduce calculation and improve the speed.At the same time,aiming at the drift problem of single target tracker,the maximum correlation response and peak sidelobe ratio are used as tracking confidence to improve the tracking performance.The performance of the optimized algorithm is improved by comparison experiments.Based on a single target tracker with great performance,studying the data association method for solving multi-target matching problem.Matching with the target feature and similarity measure,the appearance feature based on the deep learning model can extract similar features better,and the same feature extraction model structure in the single target tracker is used to reduce the calculation.In order to solve the matching problem of similar targets,the positional features are used for secondary matching to improve the accuracy of system matching.For the application scenario of traffic monitoring,the motion state of the moving vehicle is described by parameters such as target position,position prediction and confidence,and method of update and conversion is designed according to the actual scene and the different motion state of the moving vehicle.In the UA-DETRAC data set of traffic surveillance scene,the comparative experiment proves that compared with the traditional multi-target tracking algorithm which only uses the traditional feature to measure the similarity between targets,this paper updates the filter template by using the depth appearance feature combined with the tracking confidence,uses the twin structure-based appearance feature combined with the location feature twice matching,and obtains 25% and 18% on the MOTA and Prcn indices respectively.The improvement also reduces the IDswitch of vehicle tracking system,which shows great advantages in accuracy and accuracy.Based on the results of vehicle tracking,the vehicle trajectory is analyzed and processed,and different trajectory clusters are clustered by DBSCAN algorithm to realize the function of vehicle trajectory analysis.Finally,the requirements and functional modules of the vehicle tracking and trajectory analysis system are analyzed,and the vehicle tracking and trajectory analysis system based on the above algorithm is designed and implemented.