Research Of Vehicle Re-identification Algorithm Based On Depth Learning

With the development of social economy and the improvement of traffic network,the traffic volume on urban roads has increased rapidly,and road monitoring,traffic control and public transport safety are facing serious challenges.Therefore,vehicle re-identification has attracted extensive attention of researchers.Vehicle re-identification refers to the use of computer vision technology to achieve cross-camera retrieval and recognition of the same target vehicle,which can effectively improve the intelligent level of urban road traffic monitoring and management.In recent years,the vehicle re-identification algorithm based on deep learning has effectively improved the accuracy of vehicle recognition.However,due to the influence of illumination,vehicle attitude,angle of view,occlusion and other factors,large intra-class differences and high interclass similarity are still two major challenges for vehicle re-identification.To solve this problem,this paper has successively proposed a vehicle reidentification method based on multi-feature fusion and a metric learning method based on class center constraint:1.Vehicle re-identification method based on multi-feature fusion.The complex and changeable urban road traffic environment brings many uncertain factors to the feature extraction of vehicle images.It is difficult to distinguish vehicles with similar appearance only by extracting a single global feature.In order to extract vehicle features with higher discrimination and stronger stability,the method adopts a multi-branch network structure.After the main feature branch and four sub-feature branches are extracted from vehicle features,the more discriminative vehicle features are obtained through fusion.In addition,the method uses the Non-Local structure to learn the long-distance dependence between different regions of the feature layer,and uses the spatial correlation of the features of different regions of the image to improve the differentiation and stability of vehicle features.2.Measure learning method based on class center constraint.In view of the problems of weak constraint conditions,sensitive anchor selection and wrong sample movement direction in the triple loss function,this method uses neural network to fit the class center,and selects a batch of samples each time to calculate the network loss with the class center as the constraint object,in which the intra-class constraint term is used to reduce the intra-class difference,and the inter-class constraint term is used to increase the inter-class distance,Together,the two can make each sample cluster in the feature space to form a cluster.In addition,the method calculates the Jakad distance between samples based on the principle of Knearest neighbor,and then reorders and optimizes the initial search result sequence,thus improving the accuracy of vehicle re-recognition.In order to verify the effectiveness of the method,this paper carries out experimental design and algorithm performance evaluation on the common data set VeRi-776 and VehicleID for vehicle re-identification.The experimental results show that the above methods can effectively improve the effect of vehicle re-identification,and each evaluation index has been improved.