Vehicle Re-identification Research Based On Subtle Difference Recognition And Deep Learning

With the continuous promotion and development of smart city technology in our country,the scale of traffic video surveillance network continues to expand.As an vital research object in traffic surveillance video,vehicles have attracted a lot of attention in computer vision,such as vehicle detection,license plate recognition,vehicle classification and other technologies.In recent years,vehicle re-identification(Re-ID)as a sub-direction of image retrieval has gradually come into the vision of researchers,which aims to search for the same target vehicle's trajectory in the massive search database under the condition of given query target vehicle.Therefore,vehicle Re-ID has widespread applications in urban video surveillance,suspected vehicle tracking and vehicle retrieval.The existing vehicle Re-ID methods face with the following three problems.Firstly,the rigid structure of the vehicle leads to the similar appearance of the vehicle in the same year.It is necessary to extract the discernible fine-grained features to retrieve the vehicle.Secondly,since vehicle Re-ID is a cross camera image retrieval problem,the appearance of the same vehicle from different viewpoints is quite different.We need to learn the complete multi-view features and design a reasonable multi-view matching method.Thirdly,considering the expensive labor cost of large-scale vehicle labeling,it is still a very difficult work to construct large-scale vehicle Re-ID dataset.The existing methods mainly improve the robustness and cross-domain transfer ability through the data augmentation method based on style tranfer.How to design a soft label for the generated data to smooth the domain gap has become an urgent problem.This paper has conducted an in-depth research on the vehicle Re-ID task based on the above three aspects.The main research results and contributions are as follows:(1)First,aiming at the small appearance difference of the same vehicle type,and the correct sample ranks lower in the ranking list,this paper proposes a fine-grained model for vehicle Re-ID based on a two-stage re-ranking(TR).This method first presents a discriminative fine-grained network(DFNet)to learn fine-grained features with strong discrimination.The DFNet contains a siamese branch network and a fine-grained branch network.It is jointly supervised by identification loss,verification loss and fine-grained loss.Then,the extracted features of the two branch networks are fused to identify the subtle differences of vehicles.In the ranking stage of vehicle Re-ID,this paper designs a TR method to optimize the original ranking algorithm.The ranking method obtains the mean sample in the first stage of ranking,and then combines the mean sample with the ranking set of query sample to participate in the second stage retrieval.The final similarity metric combines three metric distances to retrieve more positive samples.The experimental results confirm that the proposed DFNet can accurately identify the subtle differences between vehicles in the same type,and TR can mine more positive samples from the lower rankings during the retrieval process.(2)Then,aiming at the large difference in appearance of the same vehicle from different viewpoints,and the unreliable information of the single viewpoint matching in the retrieval stage,this paper proposes a viewpoint adaptation network(VANet)with a cross-view distance metric for vehicle Re-ID.Firstly,the VANet is proposed to input the final feature vector,which fuses the feature maps from different levels to improve the capacity of viewpoint awareness.In the meantime,the VANet learns the original training data of different viewpoint pairs by style-transferred network to generate cross-view data as the data augmentationpart.A cross-view label smoothing regularization(CVLSR)is designed to assign virtual labels for the generated data.CVLSR assigns values to the corresponding classes according to the vehicle color invariance before and after style transfer.Furthermore,this paper also proposes a cascaded cross-view distance metric(CCM).In the feature extraction stage,CCM uses adaptive weight matrix to fuse the original view and cross-view features.Then,CCM uses the original view features and the fused features to learn two distance metric matrices in the ranking process,which learns the product of different weights for the two metrics as the final distance.The results on two vehicle Re-ID datasets show that the proposed method achieves better performance than others on multi-view problem.(3)Finally,aiming at the poor generalization ability of the model and the weak cross-domain transfer ability during the limited supervised learning,this paper proposes a vehicle Re-ID method based on inter-domain adaptation label smoothing regularization(IALSR).This method proposes a multi-domain joint network(MJNet).Firstly,the multi-attribute features extracted by MJNet are input into gaussian mixed model(GMM)for clustering to obtain multiple inter-domain subsets.Then,data enhancement is conducted by the style transfer between each pair of subsets.Existing label smoothing regularization assigns all the weights to one class in the generated data labels,which will lead to the over-fitting problem.Therefore,this paper designs an IALSR to preserve the self-similarity and domain-transitivity before and after translation for each generated sample.A large number of experimental results show that IALSR has stronger generalization ability for cross-domain data augmentation.