Research And Application Of Vehicle Multi-target Tracking Algorithm For Urban Road Scenes

With the improvement of people’s living standards,the number of car ownership has grown rapidly.In urban road scenes,vehicle multi-target tracking algorithms are widely used,usually as a support for applications such as vehicle behavior analysis and traffic flow statistics.However,vehicle multi-target tracking algorithms still face many challenges,which mainly include difficulty in vehicle identification,vehicle occlusion,rapid changes in vehicle speed and missed detection by vehicle detection algorithms.Based on the actual application scenarios in urban roads,this thesis conducts research on the above problems,and designs and implements a vehicle multi-target tracking algorithm.Firstly,aiming at the difficulty of vehicle identification in vehicle multi-target tracking,this thesis designs and implements a vehicle re-identification network to extract discriminative vehicle appearance features.Considering the real-time requirements of the algorithm,the backbone network part selects a relatively lightweight Res Net18 to extract appearance features.In order to improve the robustness of the vehicle re-identification model to appearance changes and occlusions,this thesis adds an IBN structure and a Drop Block regularization module to the backbone network.Aiming at the problem of background interference in vehicle images,this thesis designs and implements a lightweight hybrid attention module.The module includes two branches,channel attention and spatial attention,which filter features in channel and spatial dimensions,respectively,to enhance features related to vehicle targets and suppress irrelevant interference information.In this thesis,the model is trained and tested on the vehicle reidentification dataset Ve Ri-776,and the re-identification accuracy can reach 75.4%.Then,based on the proposed vehicle re-identification network,this thesis designs and implements a vehicle multi-target tracking algorithm.Aiming at the problem of trajectory interruption caused by the reduction of the confidence of the detection results when the vehicle is occluded,this thesis designs and implements a hierarchical data association algorithm to associate all detection results with the existing vehicle trajectories.For high-confidence detection results,use vehicle re-identification features for association.Considering that low-confidence detection results often contain unreliable appearance information,the Intersection-over-Union distance is used for association.Aiming at the problem of rapid change of vehicle speed and missed detection by vehicle detection algorithm,this thesis proposes trajectory recovery and missed detection processing strategies respectively.The trajectory recovery strategy uses vehicle reidentification features to recover interrupted vehicle trajectories;the missed detection processing strategy uses the ECO single-target tracking algorithm to compensate for missed detections.In this thesis,the evaluation dataset SP-MOT is constructed in the urban road parking lot scene.After testing,the tracking accuracy can reach 84.5%,and the speed on the RTX2080 Ti is 26.8FPS.Finally,this thesis transplants the proposed vehicle multi-target tracking algorithm to the Jetson TX2 embedded platform and applies it to the urban road scene.For the Jetson TX2 platform,Tensor RT is used to accelerate model inference,and the tracking algorithm speed is 10.3FPS.In this thesis,the tracking algorithm is applied to the urban road parking management system,and an event capture module is implemented to capture the parking and departure events in the roadside parking lot in real time.