Based On Deep Learning Research On Cross Camera Vehicle Re-Identification

With the continuous and rapid increase of motor vehicle possessing quantity,traffic congestion,traffic safety and traffic management have become increasingly serious.These problems have attracted extensive attention of scholars and society,and governments have also issued corresponding traffic development strategies.In order to strengthen road traffic management and improve traffic safety,a large number of video cameras are installed on the road.In order to enable machines to search for problem vehicles or suspected vehicles from video,vehicle re-identification has become an emerging technical hotspot in the field of intelligent transportation and road traffic safety in recent years,which has a wide range of potential roles in maintaining public traffic safety.The rapid development of deep learning has made great progress in vehicle re-identification technology,but it still faces the problems of large similarity between classes and high difference within classes.The vehicle license plate is the only identification of the vehicle,so the early or traditional vehicle re-identification task is to recognize through the license plate.However,the license plate exists the phenomenon of set plate,forgery and occlusion.In addition,the motor vehicle license plates are not captured in the video data(images)of many scenes.Therefore,this thesis uses the theory and technology of deep learning to process the video data containing vehicle images,extract the vehicle appearance features,and re recognize the vehicle according to the vehicle appearance features.The main work and innovations of this thesis are as follows:(1)In this thesis,Res Net50 network is used to extract the fine-grained features of vehicles,and triplet loss function is used to train the network model.In view of the shortcomings of the traditional triplet loss function in selecting triplet,that is,the way of randomly selecting triplet in the overall dataset,this thesis first introduces the method of batch difficult sample mining to optimize it.When selecting triplet,they are processed in batches,and the more difficult positive and negative samples are selected.So as to speed up the training speed of the network and improve the optimization of the model;Secondly,considering the relative distance between positive and negative samples and the absolute distance between positive sample pairs at the same time,the triplet loss function is further improved to ensure that the network can shorten the distance between positive sample pairs while pulling the positive and negative samples apart in the feature space.(2)Aiming at the complexity and diversity of vehicle features,a multi task vehicle re-identification method combining coarse-grained global features and fine-grained local features is proposed to overcome the problems of large intra class distance and small inter class difference in vehicle re-identification.The whole vehicle reidentification network is divided into three sub networks,two of which use Res Net18 network to extract the color and vehicle type features of vehicles,and the other uses Res Net50 network to extract the fine-grained features of vehicles.An attention mechanism is added to the Res Net50 network to extract fine-grained features to increase the robustness of the features extracted by the model.Finally,the designed vehicle re-identification network model is trained and tested on the Ve Ri dataset and Vehicle ID dataset.In the training stage,the losses of the three networks are comprehensively calculated,and the optimal solution of the training model is obtained.(3)Because this thesis uses the Multi-task learning method to extract the vehicle features,we need to consider the fusion of different losses in the training stage.In previous studies,many researchers directly add up the losses of multiple networks,which will lead to the loss of the whole network may be dominated by the loss of a certain network,and only one network has been better optimized.Some researchers add the weights of different losses through empirical method,which is highly subjective and needs to be constantly adjusted.In this thesis,the weighting method based on the homoscedastic uncertainty is introduced to fuse the different losses in the Multi-task learning network.In this way,the Bayesian uncertainty is used to model,the maximum likelihood function of Multi-task learning network is constructed,and the noise coefficient is introduced σ,Finally,the more reasonable weight of Multi-task learning network loss is obtained.Using the vehicle re-identification model designed in this thesis,experiments are carried out on Ve Ri dataset and Vehicle ID dataset respectively.The test results are compared with the results of current mainstream methods,and their m AP and Rank indexes have increased,indicating that the algorithm proposed in this thesis has higher accuracy of vehicle re-identification.