Research And Application Of Vehicle Re-Identification And Tracking Analysis In Multi-Camera

With the development of technology,traffic problems have become an important part of urban construction.Therefore,Intelligent Transportation System(ITS)came into being.ITS is designed to use the massive data processing capabilities of computers to provide solutions for evacuating traffic congestion,reducing the human and material resources in urban road supervision,and making urban traffic easier to monitor and dispatch.Vehicle tracking and re-identification in video as an important part of ITS,which has become a research hotspot in computer vision.There has been many researches on vehicle tracking in single cameras,and the current research hotspot in ITS is the problem of vehicle re-identification and tracking in multi-cameras.There are already had a lot of research awards in vehicle detection.In this thesis,after comparing the traditional vehicle detection algorithm and the vehicle detection algorithm based on machine learning,a vehicle detection algorithm based on YOLOv2 is proposed.Although there are few research awards in vehicle re-identification,with the inspiration of person re-identification,a vehicle re-identification strategy combining multiple characteristics is proposed.The vehicle detection algorithm based on YOLOv2 considers the particularity of the vehicle detection scene,aims to optimize the calculation speed and detection accuracy of YOLOv2 algorithm.In this thesis,the characteristics of vehicle detection scene which can used to optimize the YOLOv2 algorithm are analyzed and propose the improved YOLOv2 algorithm for this scene.Its effectiveness is proved by experiments.The vehicle re-identification strategy combining multiple features considers the identification informations of the vehicle re-identification,and designs the loss function by license plate,appearance,deep learning and vehicle type.Considering the importance of license plate information in vehicle re-identification sense,the license plate features are weighted to reduce the overall amount of calculation and improve the system calculation speed.The effectiveness of the vehicle re-identification system proposed in this thesis is proved by experiments.Finally,this thesis designs a multi-cameras vehicle re-identification and tracking platform.The system mainly consists of vehicle detection module and vehicle re-identification module.Both of these modules are using the vehicle detection algorithm and the vehicle re-recognition algorithm proposed in the thesis to maximize the calculation speed while ensuring the accuracy.At the same time,the multi-cameras vehicle trajectory planning module is realized.The effectiveness of the multi-camera vehicle re-identification and tracking system proposed in this thesis is proved by experiments.