Vehicle Re-identification Based On Attribute Aggregation And Multi-view Sparse Correlation Regularization Ranking

In recent years,human travel and cargo transportation have become increasingly inseparable from vehicles.The number of vehicles has increased year by year,making vehicle-related research a hot spot.The main task of vehicle re-identification is to match the video or image of the vehicle under the multi-camera network structure.The vehicle re-identification task has broad application prospects in the fields of traffic management,security monitoring,and smart city.Although the license plate is widely used as the unique identity of the vehicle,it is simple and effective to identify,but the license plate information is subject to some factors such as false license,occlusion,and low resolution.Therefore,it is important to find an approach to identify the vehicle based on the appearance of the vehicle rather than the license plate.However,due to the different properties of the camera itself,low image resolution,viewing angle changes,light changes,occlusion,and other factors,the task of vehicle re-identification based on appearance is not smooth.Besides,due to the similarity of the same model of the same brand and the great differences in the images taken by different cameras,the task of vehicle re-identification is full of challenges.Therefore,this thesis focuses on vehicle appearance,and studies and explores the image-based vehicle re-identification.The main work is as follows:(1)Considering the color and vehicle model is the most intuitive property to describe the vehicle,the deep learning framework can represent data robustly.In this work,the depth features of vehicle images are extracted by various methods.Firstly,a convolutional neural network is used to extract the feature representation of the vehicle.Secondly,because appearance is the most signification feature of vehicles,especially the color and vehicle category,which can be recognized at a glance,a multi-task deep learning framework for aggregating multi-attributes is proposed.Vehicle attributes(color and vehicle category)are added to the traditional identity classification network,that is,the loss of identity recognition and attribute recognition is added at the last layer of the network.We used two basic networks,GoogLeNet and ResNet-50,to extract multi-view features,and then fuse these two features to represent vehicle information synergistically.Finally,a large number of experiments are carried out on the open vehicle re-identification datasets VeRi-776 and VehicleID to demonstrate the superiority of the method,and verify the promotion and contribution of the proposed aggregated multi-attribute multi-task deep learning framework to vehicle re-identification.(2)Due to the insufficient information contained in a single perspective,it is impossible to obtain the depth features of the vehicle image in all directions.To compensate for the deviation caused by a single perspective and to mine the consistency between multi-view features,this work proposes a robust multi-view feature.The related sparse sorting model of sticks introduces the consistency constraint of multi-view weights and related items between views in the multi-view sparse representation model.In particular,sparse coding is a weighted linear combination that approximates an input vector as a small number of basis vectors from a dictionary,while minimizing the diversity between sparse coefficients of any two views to explore the correlation of multi-view features.After optimizing the model,the sparse coefficients of each view are added as the final similarity ranking result.At the same time,the method can also be viewed as a general framework for multi-view feature fusion of any existing network.It can be seen from the experimental results that the vehicle re-identification method based on multi-view correlation sparse sorting proposed in this chapter performs better on the VeRi-776 dataset.Besides,since the initial ranking directly compares the distances between two images,the correlation between similar images is neglected.Finally,an effective post-processing method,reordering technique,is used to reorder the initial search results to improve the retrieval accuracy.For each pair of images,we only accumulate the distance between the immediate neighborhood of each image and another image,which leads to the reordering of promising images.The method is fully automated,unsupervised,and does not need to calculate a new ranking list.The experimental results show that the model achieves recognition accuracy of 90.5%on the VeRi-776 dataset.