Research On Vehicle Re-identification In Non-Overlapping Domain

Traffic surveillance has a big impact on the current public transportation system,which is mainly used for traffic control and safety.The vehicle is an important target in the urban surveillance system.Over the past decade,vehicle-related research has gradually become a hot topic and have made great progress,such as vehicle detection,classification,and segmentation.Unlike these tasks,vehicle re-identification aims to accurately match a specific vehicle in samples captured by multiple cameras.With the vehicle re-identification algorithm,the matching work can be performed automatically,which not only saves labor but also takes less time.Hence,vehicle re-identification plays a vital role in building smart cities and safe cities.With the development of deep learning,compared with the method of using traditional features for image matching,the deep features extracted by convolutional neural networks for training re-identification model has become a trend.However,these methods still faces some problems.In this thesis,three key problems are focused:one is how to extract effective discriminative features,the second is how to improve the generalization ability of the model,the third is how to solve the model training problem in the absence of annotation.Therefore,this thesis proposes the following:Aiming at the problem of extracting effective features,this thesis proposes a multi-region aware model with strong discriminative features for re-identification task.In the model,in addition to extracting global features,detailed features are extracted from a series of local regions as local features.Different from extracting local features by directly dividing rigid features,this thesis introduces a localization model,which is based on spatial transformation network to exploit local regions with more unique visual cues.In addition,in order to further improve the accuracy of the vehicle re-identification task,a context-based ranking method is designed to generate a secondary ranking list by considering the similarity between neighbors.In order to verify the effectiveness of the proposed method,this thesis designs a large number of experiments,which show that our method has improvements on different datasets.For cross-domain problems,especially there is an evident drop of accuracy when the re-identification model trained on one domain while tested on another one.This thesis proposes a domain adaptive framework for vehicle re-identification task.To smooth the cross-domain bias and make full use of the existing well-labeled samples,our method utilizes a two-branch adversarial network for image-to-image translation,which attempts to make the labeled images in the source domain learn the style of unlabeled images in the target domain.Especially,there isn't any annotations during the translation.Besides that,the identity information of the source domain can be preserved after the translation.In order to better suppress background information and focus on foreground information,this thesis proposes an attention-based feature learning network to train the vehicle re-identification model.Through our proposed method,the well-trained re-identification model has better generalization ability for various scenes in reality.To verify the effectivenss of the proposed method,some experiments are conducted on existing datasets.The experimental results show that our method could smooth the domain bias and improve the generalization ability of the re-identification model.Due to lacking of annotations,the vehicle re-identification model couldn't be fine-tune in the target domain directly,which leads to a bad performance when testing the well-trained model in the target domain directly.Hence,this thesis proposes a novel progressive adaptive learning method for vehicle re-identification which aims to infer the relations from large amounts of data without annotations.In our method,in order to speed up the convergence rate of the model,the adversarial generation network is utilized to generate images,which has similar data distribution with the unlabeled target domain as the "pseudo target sample".These pseudo samples are combined with the unlabeled samples selected by the dynamic sampling strategy to jointly train the vehicle re-identification model in the subsequent stage.In addition,due to the inaccurate of the clustering results,this thesis further proposes the weighted label smoothing loss,which considers the similarity between samples with different clusters to balance the confidence of pseudo labels.Through iterative and alternating training model and clustering,the re-identification model gradually adapts to the target domain.From the experimental results,as the number of iterations increases,it also could be seen that the re-identification model has a certain improvementsFor the cross-domain problem,a multi-label learning approach is proposed to exploit the potential similarity of unlabeled samples to build multiple clusters from different views automatically.Different from existing methods,our method takes the multi-view into account.For one vehicle,it could describe it in different views,such as the color and the type.Therefore,our adaptive multi-label learning employs the potential similarity of unlabeled samples.And multiple features could be generated by the proposed focus drop network,which are utilized to automatically construct multiple clusters from different views.Specially,for each vehicle in the target domain,several different annotations can be assigned according clustering,which are employed to train the vehicle re-identification model.The clustering and training process are conducted repeatedly and alternatively until the re-identification model is stable.In addition,due to the inaccuracy of the clustering results,different from the triplet loss,the hard triplet center loss is proposed to take the difference of intra-cluster and inter-cluster into consideration to better training the unsupervised framework to adapt the unknown domain.Finally,to verify the effectiveness of the proposed method,this thesis does various experiments on several datasets that are constructed with a large amount of vehicle images from surveillance videos.And our methods get the competitive results with some existing methods.