Research On Vehicle Re-identification Based On Multilevel And Multi-branch Network

With the continuous development and progress of society,the living standards of the people are increasing day by day.As an important means of transportation,the number of vehicles is increasing year by year,which brings great challenges to traffic management.Whether it is possible to accurately search for a specific vehicle has become an urgent problem to be solved in the field of public safety,so the task of vehicle re-identification has received more and more attention.The purpose of vehicle re-identification is to use the computer to judge whether the vehicles emerging in different cameras are the same vehicle,thereby effectively reducing the work pressure of the relevant staff and improving the efficiency of case investigation.Since the quality of the vehicle images captured by the monitoring system under actual scenarios cannot be guaranteed,it is difficult to obtain license plate information in it,and some vehicles have false licenses and decks.Therefore,it is necessary to quickly determine the target vehicle and use vehicle feature information.To solve this problem,this paper designs a corresponding vehicle feature extraction network structure based on deep learning.The research content of this article is as follows:Firstly,based on the Convolutional Neural Network(CNN),this paper proposes a multi-branch vehicle feature extraction network structure.By adapting the random occlusion mode of the vehicle characteristics,the local feature strengthening branch is completed in conjunction with the corresponding network structure design,and the overall feature extraction branch is combined to form a multi-branch vehicle feature extraction network,which obtains the "discriminatory" vehicle features.Secondly,based on the vehicle feature extraction structure,through the analysis of the advantages and disadvantages of the loss function that is used more frequently in the vehicle re-identification task,this paper introduces the center loss function into it,and completes the design of the loss function in conjunction with the cross entropy loss function.The effectiveness of the loss function and vehicle feature extraction network structure designed in this paper is verified through experiments.Finally,since most of the vehicle feature extraction networks are deep,it is easy to have problems such as the gradient update of the front layer network is less affected by the back layer network,the utilization of key feature information is smaller,and the single classification model has limitations.In this paper,a multi-stage series model is proposed,and the network structure proposed in this paper is optimized based on this model,and the overall algorithm design of the multi-level and multi-branch vehicle reidentification network in this paper is completed.The algorithm proposed in this paper is tested on multiple public data sets,and the test results verify its rationality and effectiveness.