Person Re-Identification Based On Graph Matching And Structure Learning

In the field of computer vision,person re-identification plays an important role.The main task of person re-identification is to conduct cross-camera person target recognition by computing the similarity between different person images.This technique can be widely used in many applications in both transportation and security area.It is known that person re-identification still faces many challenges,such as background clutter,lighting variation,low resolution,and dramatic changes in the scene.Among them,the problem of visual misalignment caused by changes in the appearance of human poses and camera views is one of the main challenges in person re-identification tasks.In recent years,research in this area has focused on two aspects,i.e.,1)designing appropriate features to represent human' appearance and 2)finding suitable metrics to measure the similarity between images across cameras.However,most of these methods can not be used to address the problem of spatial misalignment between local patches existing in the image pair.While some patch-based methods are effective in solving this problem,this thesis addresses the challenges in the mentioned person re-identification by using patch-based matching methods,as follows:(1)In order to suppress the influence of outliers in patch-based matching,this thesis proposes a person re-identification method by using a local sparse matching model.First,we divide the image into overlapping patches and use these patches as graph nodes to construct a graph.The relationship among them is the edge of the graph.A local sparse matching model is used to obtain a more robust patch-based correspondence of positive sample pairs.Second,based on the similarity comparison,we can obtain the patch-based correspondence between few pairs of examples that are most similarly to the test pair.Finally,we apply the correspondence between the learned with test image pair and use the local global metric method to obtain the final re-identification result.Using the patch-based local sparse matching algorithm we can obtain more robust correspondences,and the local global distancemeasurements also make the final result more reliable.Experimental results on several datasets show that the proposed method can achieve better recognition of pedestrians with different poses.(2)For camera angle,there are certain perspective rules to follow for person images taken from a fixed camera angle.Based on this,we propose to group images under different cameras by pose,and obtain the consistent structure correspondence between pose groups.That is,the correspondence between poses is obtained by using the proposed multi-graph cooperative matching model.During the testing stage,the pose pairs corresponding to the test pair is found through pose parse,and the correspondence structure of the pose group is used for correspondence structure learning.In this way,we can get closer to the realistic scene.Different from the traditional graph matching problem,the multi-graph cooperative matching problem aims to mine the matching relationship between different image group,and learn the correspondence between patches under different perspectives.The experimental results show that it can obtains better recognition performance for person re-identification tasks under pose changes.