Research On Vehicle Re-Identification Method Based On Deep Learning

Vehicle re-identification refers to a given image of a vehicle,which retrieves all images of the same vehicle taken by different cameras in the vehicle image library based on the appearance information of the vehicle.Because the individual appearance of different vehicles of the same type is very similar,and there is obvious problem of similarity between classes,it is necessary to distinguish different vehicles by subtle differences in their appearance.In the current research on vehicle re-identification technology,there are still problems such as low efficiency and real-time performance due to the large number of parameters in the vehicle reidentification model,and low recognition accuracy due to the lack of global understanding of the image in the vehicle re-identification model.To solve these problems,this paper presents a vehicle re-identification method based on multi-granularity feature and multi-feature fusion training and a lightweight vehicle re-identification method based on Transformer.The main research work in this paper is as follows:(1)To solve the problem of similarity between classes in vehicle re-identification,this paper presents a vehicle re-identification method based on multi-granularity feature and multifeature fusion training.This method first extracts the multi-granularity features of vehicle images using a dual-branch network,extracts local distinguishing features and global features with global semantic information from different dimensions,thereby improving the ability of the model to learn more discriminant features.Secondly,a feature fusion method for multifeature fusion training is proposed,which combines multiple features together to preserve the key features effectively,thus improving the robustness of the model.The proposed method is validated on several open datasets.The experimental results show that the proposed method achieves 80.10% and 72.32% m AP and 96.25% and 62.16% Rank-1 respectively on Ve Ri-776 and Vehicle ID(Large)test sets.(2)This paper presents a lightweight vehicle re-identification method based on Transformer to solve the problems of too many parameters of vehicle re-identification models,lack of global understanding of images,and inability to model dependencies between features.First,the method uses the Shuffle V2 module and the Shuffle Vi T module proposed in this paper to extract the features of the vehicle image,so that the features effectively preserve the spatial relationship between regions,and uses the Transformer’s self-attention mechanism to model the dependencies between features to enhance the global understanding of the image.Secondly,the method mixes Sup Con loss function with cross-entropy loss function to enhance the constraints between the same vehicle image features,so as to improve the discriminant ability of the features.This paper designs and performs comparative and ablation experiments.The experimental results on Ve Ri-776 and Vehicle ID datasets demonstrate the validity and feasibility of the proposed method.(3)A road intelligent monitoring and anomaly detection system is designed and implemented.According to the characteristics of the scene of vehicle tracking suspected of violent terror,the system architecture is designed and the related modules are implemented.To solve the problem that the abnormal vehicle detection module does not have a special abnormal vehicle detection dataset,this paper collects the abnormal vehicle images and labels an abnormal vehicle detection dataset.Finally,the system realizes the designed function,and the results of function demonstration show that the system can meet the needs of abnormal vehicle re-identification and tracking.