Research On Cross-view Person Re-identification

With the promotion of the Safe City and Smart City programs,security video surveillance system has been widely used in most populated cities in China.This system plays an increasingly important role in fighting criminals,improving the management of urban public safety,and safeguarding the social stability.It not only should satisfy the basic requirements,such as videos collection,transmission,storage and displaying,but also need to be capable of analyzing and processing the data intelligently.It has become the focus of most researchers to intelligently analyze and process the collected data.In recent years,person re-identification,as an intelligent video analysis technique in SVSS,has attracted increasing attention in computer vision.It aims to recognize person-of-interest from non-overlapping camera-views.This technique can facilitate pedestrian retrieval in large-scale surveillance network.To tackle this problem,massive efforts have been made.However,it remains a challenging problem since a person’s appearance often undergoes dramatic variations across camera views due to changes in view angle,body pose,illumination and background clutter.How to build a more effective person image matching model is the main problem to be solved.This dissertation aims to improve the person re-identification by the following efforts.1)Learning a suitable distance metric is the key problem in person re-identification.Most existing metric learning methods usually adopt the global decision rule,which may suffer from sub-optimal learning performance when coping with some real-world tasks with more complex inter-class and intra-class variations.To alleviate thus limitation,this dissertation presents a locally adaptive decision rule for metric learning.We jointly learn two metrics in both original feature space and auxiliary feature(privileged feature)space.Given a pair of person images,the distance in privileged feature space functions as a local decision threshold for the guidance of metric learning in original feature space.The distance metric learned in original feature space is used for distance computing in testing set.The proposed metric learning method has been evaluated in multiple person re-identification datasets.Experimental results have demonstrated that the proposed method can obtain better performance than global decision threshold based methods.2)Most existing person re-identification approaches are fully-supervised,requiring access to a large amount of labeled training image pairs.However,labeling data is very costly and time-consuming.For this reason,this dissertation proposes to enhance person re-identification in a self-trained subspace in a semi-supervised setting.It can exploit both labeled and unlabeled data to learn a low-dimensional and discriminative subspace.We first utilize the labeled data to learn an initial discriminative projection,which further projects all the unlabeled data into a low-dimensional subspace.Then,we construct pseudo pairwise relationships among the unlabeled persons using k-nearest-neighbors algorithm in this low-dimensional subspace.which is further combined with the labeled data to learn a discriminative projection.Given the newly learned projection,we refine the pseudo pairwise relationships and relearned the discriminative projection with these updated pseudo pairwise relationships.This process is iterated until the pseudo pairwise relationships remain unchanged.In this way,the discriminative ability of the learned subspace can be significantly enhanced.Extensive experiments on multiple person re-identification datasets have showed that the proposed semi-supervised approach can improve the fully-supervised approach(using labeled data only)by a large ’margin.3)Traditional person re-identification approaches consist of two separate steps:feature extraction and metric learning.Once useful information has been lost in the feature extraction stage,it can hardly be recovered later.This dissertation develops a novel deep metric learning scheme by jointly learning feature representation and distance metric in an end-to-end manner.The main contribution is the development of a structural deep metric learning objective function.In the proposed learning objective,each positive pair is allowed to be compared against with all negative pairs.To better leverage the hard positive samples,each positive pair is adaptively assigned a hardness-aware weight in the proposed learning objective to modulate its contribution to the network training.Moreover,the second-order statistic of positive pair distances is employed to construct a global loss term,which is further incorporated into the proposed structural metric learning objective to improve the generalization capability of the network.Extensive evaluations.on three large scale person re-identification datasets have demonstrated that the proposed method can improve the baseline method effectively.In summary,to alleviate the limitation in existing person re-identification methods,this dissertation has introduced a locally adaptive metric learning method,a semi-supervised subspace learning method based on self-training,and a structural deep metric learning method,respectively.Extensive experimental results have verified the effectiveness of the proposed three methods.