Research On Vehicle Tracking Algorithm Based On Deep Learning For Roadside Camera

Intelligent transportation system is a comprehensive transportation management system integrating communication technology,transportation infrastructure,vehicles and users,which is of great significance in alleviating traffic congestion,improving driving comfort and increasing road safety.As the basis and premise for realizing other advanced functions of the intelligent transportation system,environmental perception is composed of vehicle tracking from the roadside perspective as the core technology.By obtaining the trajectory information of multiple vehicles,it provides all-round support for vehicle operation and supervision,and effectively improves the road traffic rate.In this paper,a vehicle tracking algorithm based on road side view was proposed to meet the requirements of obtaining vehicle position number and trajectory information from the vision sensor deployed on the road side.The main research contents are as follows:(1)Vehicle tracking dataset from roadside perspective was constructed.This paper compared and analyzed the mainstream public datasets of vehicle tracking,and selected UA-DETRAC as the basic training and testing set.In order to improve the coverage of the dataset to the realistic scene,cameras were used to collect and annotate traffic flow videos of different roadside perspectives,different time periods and rain and snow conditions.A total of 7,328 images were collected,which enriched the diversity of perspective and weather of the vehicle target training samples in this paper.(2)A Fair MOT-V vehicle tracking model suitable for roadside perspective was built.In this paper,based on the multiple object tracking model Fair MOT,aiming at the appearance deformation caused by rapid vehicle driving and the poor classification accuracy caused by vehicle appearance similarity,the backbone network was optimized to improve the feature extraction ability of vehicle class targets,and the depth cosine measurement learning was optimized to improve the classification accuracy.The ablation test verified that the tracking accuracy of Fair MOT-V model reached 78.9%,1.2% higher than that of baseline model,and good vehicle trajectory information was obtained.(3)Optimized the data association strategy applicable to occlusion scenes.In order to solve the problems such as ID switch and trajectory fragmentation caused by vehicles blocking each other in traffic scenes,an intra-frame nearest neighbor matching association strategy was proposed to improve the success rate of target and trajectory association.In order to improve the accuracy of vehicle motion estimation,the detection confidence was used to judge the occlusion and the optical flow branch was added to predict the vehicle position information.The ablation test verified that the Fair MOT-V2 model reduced the number of ID jumps by 25.5% compared with the baseline model,and improved the tracking accuracy by 0.4% compared with the Fair MOT-V model,further improving the robustness of the model.(4)Roadside perspective vehicle tracking test for Fair MOT-V2 model was conducted.The operation interface was developed based on ROS platform to realize the visual encapsulation of the model.Based on the roadside sensing platform,the research group conducted model performance tests in three scenarios to verify the feasibility of Fair MOT-V2.The results showed that Fair MOT-V2 model can complete the vehicle tracking task well in different traffic scenarios,which has high research significance.