Algorithm And Application Of Vehicle Re-identification Based On Key Points Mask

Cross-camera vehicle identity authentication has always been an important part of applications such as suspect tracking and traffic accident liability determination.Although there exist many vehicle identity authentication systems that use license plate recognition as the main method,it is often difficult for surveillance cameras to effectively capture license plates due to problems such as dense traffic,and low shooting quality.On the other hand,the identity information brought by the license plate is not completely credible owing to the fake license plate.When the license plate cannot be effectively recognized,the visual information of the whole vehicle body becomes the key to the cross-camera vehicle identity authentication.Therefore,this dissertation studies the problem of cross-camera vehicle authentication based on the overall vehicle information and low data quality.The problem is also called vehicle reidentification problem.The research work of this dissertation is mainly on the analysis and improvement of the vehicle re-identification problem.The main contributions are as follows:(1)First,this dissertation designs a vehicle re-identification algorithm based on key points mask and multi-task learning of identity,type and color.The algorithm suppresses the response of the neural network model to the visually significant areas such as key points through key points mask,and thus enhances the focus on the nonsignificant areas of the vehicle body.The algorithm can perform type and color classification tasks while completing identity authentication.Based on type and color information,the accuracy of vehicle re-identification is further improved.(2)Second,this dissertation discusses the utilization of vehicle spatio-temporal information in practical applications.And this dissertation designs two utilization methods: spatio-temporal information filtering secondary sorting and path reachability secondary sorting.The former excludes the vehicle records outside the single movement area based on the limitations of vehicle movement.The latter uses an algorithm to generate vehicle likelihood paths and sorts the gallery records to be matched based on these generated likelihood paths.Finally,this dissertation implements a vehicle re-identification system,which is based on the vehicle re-identification algorithm.The system runs on a sandbox container.While completing the vehicle re-identification task,it can provide a suspected travel path which is based on vehicle images and spatiotemporal information.