Research On Expressway Vehicle Recognition And Tracking Technology Based On Deep Learning

One of the core technologies in the development of intelligent transportation systems is vehicle recognition and tracking on expressway,which enables real-time monitoring and management of vehicles.The use of deep learning methods for accurate tracking aims to obtain real-time vehicle driving status and trajectory information,providing important data support and decision-making basis for traffic management departments.Based on the requirements of vehicle recognition and tracking,the thesis divides deep learning-based highway vehicle recognition and tracking into three sub-tasks: vehicle detection,vehicle tracking,and vehicle re-identification,and conducts research on abnormal behavior recognition for tracking trajectories.The main research content of thesis is as follows:(1)For the vehicle detection task,the thesis addressed the complex research scenarios and image interference in monitoring by constructing a highway monitoring dataset.Based on the YOLOv5 algorithm as the basic model,the thesis applied a variety of optimization techniques to improve the model’s detection accuracy.The optimization techniques include improving the prior box clustering algorithm,using soft non-maximum suppression,using a bidirectional feature pyramid network structure with cross-layer connections,introducing convolutional block attention modules,and adding a small-scale object detection head.After optimization,the mean average precision(mAP)of the vehicle detection model on our self-built dataset reached83.5%,which was 4.5% higher than the unoptimized model.This fully demonstrates the powerful ability of the optimized model in completing the vehicle detection task.(2)For the vehicle tracking task,to meet the real-time and accuracy requirements of vehicle tracking,the thesis built a multi-vehicle tracking model based on the DeepSORT tracking algorithm.Then,the thesis combined it with the vehicle detection model to improve the tracking speed and accuracy of the model,and solve the problem of vehicle tracking box jitter caused by slow vehicle detection speed.the thesis retrained the vehicle appearance feature extraction network to make the deep appearance descriptor more in line with vehicle features,and optimized the model parameters to improve the details of the vehicle tracking model according to actual scene requirements.Ultimately,the improved model achieved an accuracy of 70.6%and a precision of 81.0%.(3)For the vehicle re-identification task,the thesis addressed the problem of difficult and inaccurate vehicle feature extraction by using attribute features to optimize image features.Based on a multi-task learning framework,the thesis built a vehicle re-identification model,and introduced an attention module in the model to generate representative attribute features and reweighted the attribute features to compensate for the global features.The re-identification model was trained and validated on the VeRi-776 dataset and Vehicle-ID dataset,and was applied to research scenarios.Ultimately,the cumulative matching curve of the model showed better recognition performance compared to other models,and had strong robustness,with a Rank@10 metric of 72.8% for queries.Based on the vehicle tracking implementation in the monitoring perspective described above,the thesis used coordinate calibration to obtain the vehicle’s motion trajectory.Then,based on the vehicle re-identification results,the trajectories of the same vehicle were applied to the stitching of vehicle trajectories under adjacent cameras and the prediction of blind area trajectories.Subsequently,a set of trajectory-based vehicle behavior detection models was established,mainly aimed at detecting behaviors such as retrograde driving,parking,speeding,slow driving,and dangerous lane changes,and applied to practical research scenarios.