Research On Person Re-identification For Intelligent Video Surveillance

In recent years,the rapid expansion of camera networks has provided massive amounts of visual data for intelligent video surveillance,but it also greatly increased the difficulty of video processing.By identifying and searching a specific person over large scale camera network,person re-identification can effectively associate people across camera views,thus plays an important role in many surveillance applications,such as cross-camera object tracking and behavior analysis.However,person re-identification is a very challenging problem since the appearance of a person can vary significantly due to cross-camera changes in viewpoint,pose and illumination.Based on the extensive review of the state-of-the-art methods,this thesis focuses on the study of appearance description and distance metric learning in person re-identification.A series of original solutions are proposed to overcome the impact of cross-camera variations and achieve encouraging performance.The main content and contributions of this thesis are summarized as follows:1.A new person re-identification algorithm based on salient region learning is proposed to address the problem that most appearance description methods ignore the diversity of human body regions.By jointly considering the visual and spatial relationships among over-segmented regions,the proposed method can learn distinct regions which follow certain spatial constraints and are insensitive to appearance changes.Meanwhile,the salieny learning is formulated as an unsupervised learning problem based on hypergraph and needs only few images of a same person as input,thus offering a much more practical solution than supervised appearance description algorithms.Experimental results show that the proposed approach can achieve better performance than existing appearance-based person re-identification methods.2.An adaptive metric learning(AML)model is proposed to cope with the inconsistent feature distributions across camera views caused by cross-camera variations.Different from conventional metric learning methods which treat all the training samples equally,AML adaptively classifies the training samples into different groups and pays different attention to them.By fully exploiting the discriminative information among training samples,AML can generate a more effective metric for person.re-identification.Moreover,the optimization of AML is formulated as a smooth convex problem,which can be solved effeicently.Experimental results demonstrate the advantage of the proposed model in terms of both effectiveness and efficiency.3.A novel framework for person re-identification is proposed to make full use of the intrinsic information existing between intra-camera images.First,the training model forces the images of the same person to have stronger affinity in the learned feature space by considering both the inter-camera and intra-camera constraints during metric learing.Then,the testing phase of existing metric learning approaches is also improved.By employing probabilistic hypergraph model to capture the neighborhood relationships among testing samples,the proposed testing strategy can provide more robust and effective ranking results than conventional pairwise matching.Experiments conducted on five datasets clearly demonstrate the superiority of the proposed framework over existing metric learning based person re-identification methods.4.To overcome the problem that using one specific metric for person re-identification often suffers from over-fitting and may not be sufficient enough to cope with all kinds of cross-camera variations,a powerful metric fusion method is proposed to combine multiple given distance metrics.The proposed method represents given metrics as different graphs and then formulates the fusion problem as a multi-graph constrained transductive learning problem.In this way,the complementary information provided by different input metrics can be effectively integrated.Furthermore,by introducing a weight learning strategy,the fusion weights of different metrics can be adaptively modulated.Experimental results show that the proposed method brings remarkable performance improvements to individual metric learning algorithms and outperforms the state-of-the-arts by a large margin.