Vehicle Re-identification Based On Multi-view Sparse Ranking And Multi-scale Attention Mechanism

As the important industrial distribution of the 13 th Five-Year Plan,intelligent transportation has risen to the strategic task of developing the information technology.For this reason,tens of millions of video surveillance terminals have been deployed nationwide.How to fully explore and utilize the traffic visual data has become an urgent problem to be solved in the construction of intelligent transportation in China.The task of vehicle re-identification is to match the vehicle image with same identity captured by the non-overlapping surveillance cameras,and ultimately obtain the vehicle trajectory in the whole camera network.Therefore,the research of vehicle re-identification can alleviate the contradiction between massive data and insufficient labor in some extent.Therefore the research plays an important role in the field of intelligent transportation,and has important application value in the development of public security,intelligent monitoring,intelligent transportation,driverless and other industries.However,vehicle re-identification has great challenge on the way of applying:(1)Due to the complex monitoring scenarios,the differences in the attributes of monitoring cameras,the change of illumination and viewpoint,there are big difference between vehicle images with same vehicle ID.Therefore it is essential to learn robust vehicle features,which can reduce the affect of illumination changes,viewpoint differences and other factors.(2)Since the vehicle is a rigid body structure,different vehicles often have very similar appearances.Therefore it is essential to distinguish the similar vehicles with fine detail differences(1)In order to reduce the effect of illumination changes,viewpoint differences and other factors to obtain discriminative vehicle features,this thesis proposes a vehicle re-identification framework based on sparse ranking to fuse multi-view information.This model enforces multiview correlation to explore the correlations of multi-view features and a Laplacian regularization as a smooth operator to preserve the local manifold structure.Specifically,we first employ multiple baseline networks to generate multi-view features.Then we explore the feature correlation via enforcing the correlation term into the multi-view Laplacian sparse ranking framework.The original rankings are obtained by the reconstruction coefficients between probe and gallery.Finally,we utilize a re-ranking technique to further boost the performance.The method effectively fuses multi-view information to obtain final feature,which is more robust for vehicle re-identification.(2)In order to mine discriminative cues to solve the challenge of different vehicles with similar appearance,this thesis designs an end-to-end multi-scale attentive vehicle re-identification network.This network fuses complementary information from multiple scale and exploits attention mechanism to mine discriminative local cues.Specifically,we feed the input image into the backbone network,and then exploit bilinear interpolation to generate multiple scale feature maps.Each scale feature map is then fed into corresponding subnetwork,followed by the spatial-channel attention block.After training these subnetworks,we use the embedding layers to fuse the multi-scale attentive feature maps and fine-tune the whole network.This network combines multi-scale mechanism and attention mechanism to obtain more discriminative features,which can boost the performance of vehicle re-identification especially for the vehicles with similar appearance.