Combination Of Appearance And License Plate Features For Vehicle Re-identification

Vehicle re-identification,as an important part of computer vision and re-identification,has attracted more and more researchers’ attention in recent years.The main reason is that more and more people in China own cars,so the challenges faced by traffic management departments are increasing.Therefore,the government has issued a series of related policies to promote the development of smart cities and intelligent transportation,which makes more and more researchers pay attention to vehicle re-identification.Thanks to years of research on face recognition and person re-identification,many methods on those fields have significant influence on vehicle re-identification.However,due to the lack of texture information and dramatic changes in the attitude of vehicles,coupled with a large number of vehicles with similar or identical models,vehicle re-identification is more challenging than face recognition and person re-identification.To this end,this paper proposes a two-module algorithm based on the appearance and license plate features of vehicle images,which makes full use of the appearance and license plate information of the vehicle.The main research contents of this paper are as follows:(1)Analyze the different characteristics of vehicle re-identification and person reidentification and point out the difficulties and particularities of vehicle re-identification relative to person re-identification.The difficulties mainly lies in the fact that vehicle images carry less information that can uniquely identify them than person images.The special points are mainly two aspects.On the one hand,compared with pedestrian images,the inconsistency problem of vehicle images is more serious,which leads to the issue that the extraction of local features from the vehicle images will reduce the accuracy of vehicle re-identification.To this end,this paper proposes a Two-Branch Network focused on extracting global features,which inspired from a person re-identification network named MGN.It deletes the operation in MGN that first spit the feature map into multiple grids then pool these grids separately.On the other hand,the vehicle contains license plate,which can be used as the unique identification of the vehicle,so the feature of the license plate can be used as the auxiliary information of vehicle re-identification.(2)Since the model based on the global feature is easy to overlook the information that is less significant or of low frequency in the images,this paper designs a Focal Contrastive Loss(FCL)function,trying to make the network can excavate difficult samples in the training process.Therefore,the network can be able to capture less significant information in the images.The FCL,which is inspired from Focal Loss,a loss function in object detection,can adaptively assign more weight to difficult samples according to its own mathematical characteristics.Due to the fact that the number of sample pairs is expanded in the contrastive loss,the sample utilization rate of contrastive loss is low.On the contrast,the sample utilization rate of Softmax Loss is high.Therefore,this paper uses Softmax Loss and Focal Contrastive Loss to optimize the Two-Branch Network proposed in this paper jointly.(3)Considering the importance of license plate information for vehicle re-identification,this paper designs a license plate re-identification network based on the bidirectional LSTM unit to extract the feature of license plate images.Finally,this paper combines the license plate re-identification network and the Two-Branch Network jointly optimized by Softmax Loss and Focal Contrastive Loss to form a two-module algorithm that includes the appearance module and the license plate module of vehicle images,which uses both appearance and license plate information of the vehicle for the vehicle re-identification task.