Research On Vehicle Tracking Algorithm In Intelligent Transportation System

Vehicle tracking is an important unit in the field of target tracking.Nowadays,computer vision and target tracking are becoming more and more popular.Due to the uniqueness of target and the wide application in real life,vehicle tracking have attracted the attention of scholars and engineers all over the world.In intelligent transportation system,vehicle tracking is a particularly important research task.The apparent characteristics of vehicle targets,complex interference in traffic scenes,camera movement and video quality all bring a lot of uncertainties and great challenges to vehicle tracking.Vehicle tracking can optimize intelligent transportation system,which is of great significance to vehicle management,road traffic design and even urban planning.On the basis of analyzing the principle of vehicle tracking and summarizing the research results,this paper completes the following work in order to improve the accuracy and robustness of vehicle tracking algorithm.1)In the target tracking system,the detection accuracy will affect the tracking effect.This chapter aiming at the problem of low detection accuracy caused by poor information extraction ability of the detection algorithm.This measure improves the deep learning backbone network framework Darknet53 and proposes the C-YOLO detection method.The algorithm adds residual blocks to change the feature extraction depth of the convolutional layer.The minimum feature graph is changed from 13×13 to 8×8,and the sensitivity field of the feature graph is increased,so that the detection method can better deal with the problem of vehicle apparent feature change in the traffic scene.The experimental results show that the algorithm has certain processing ability for vehicle occlusion,small-size vehicle target detection,background clutter interference and other problems,and the detection accuracy is effectively improved.2)Aiming at the tracking drift of KCF nuclear correlation filter algorithm in the face of interference,a Y-KCF vehicle tracking algorithm based on Darknet53 network and nuclear correlation filter was proposed.Based on discriminative ideas,the algorithm uses the Darknet53 detection framework of deep learning to assist in correcting the tracking results,then update the filter with the corrected results.Thereby providing a template for subsequent tracking process.Experiments show that this algorithm solves the problem of poor results of the correlation filter tracking algorithm when the vehicle is interfered and occluded.Especially when the vehicle target is scaled,the tracking result is significantly improved,the real-time performance is better and meets actual needs.3)Aiming at the problem of fuzzy vehicle information and single category information in multi-target tracking,an improved tracking method is proposed.The algorithm takes advantage of the network framework of deep learning to improve the detection module in SORT,and draws on the idea of tracking-by-detection.First,YOLOv3 detection method is used to extract the coordinates and category information of all vehicle targets in the video frame,which are provided to the Kalman Filter tracker to obtain the tracking frame of each vehicle.Then the Hungarian matching algorithm is used to cascade the detection results and the tracking results,determine the degree of matching according to the threshold,and distinguish different vehicle targets to complete the tracking task.Finally,the filter template is updated for the subsequent tracking.Experimental results show that the algorithm can adapt to multi-target vehicle tracking in road traffic scenes and has good performance.