Person Re-Identification Based On Visual Saliency Enhancement And Structured Low-Rank And Sparse Learning

The problem of person re-identification(Re-ID)based on the monitoring scene is a hot topic in the field of computer vision at present.The main task is to match the person in the camera networks in the surveillance scene.It has wide applications especially in the multi-camera network structure with non-overlapping views such as smart city,security monitoring and traffic management.However,it faces big challenges due to the various changes including lighting,background,pedestrian appearance and camera views caused by the non-overlapping structure of the multi-camera network.The research directions of person re-identification mainly fail into two categories:(1)Appearance based methods to leverage the illumination,pose and viewpoint changes in Re-ID by appearance modeling.(2)Learning based methods to bridge the appearance gaps between the low-level feature and high-level human semantic.From the data point of view,Re-ID problem can be divided into(1)Single-shot Re-ID,where only a single frame/image is recorded for each person within each camera view.(2)Multi-shot Re-ID,where a video sequence is recorded for each person.Based on these two data morphologies,the main work of this thesis can be summarized as follows:(1)For single-shot Re-ID,based on the fact that person images are generally the bounding boxes that consist of unnecessary background or occlusions,we propose a novel saliency weighted feature descriptor via graph based manifold ranking.Specifically,a close-loop graph is constructed on the superpixel nodes,which is further ranked by the weights based on their distance to the background and foreground cues.We further integrate the superpixel saliency into a patch based feature descriptor,local maximal occurrence,to construct the saliency weighted feature descriptor for person re-identification,which can enhance the person area and suppress the background or occlusions.(2)For multi-shot Re-ID,we assume that the appearances of those subset images with similar viewpoint against camera draw from the same low-rank subspace,and all the images of a person under a camera lie on a union of low-rank subspaces.Based on this assumption,we an effective subspace learning approach for multi-shot Re-ID in the low-rank representation frame-work,in which the nonnegative,low-rank and sparse constraints are simultaneously employed to construct an informative graph for refining the affinities among person images.Moreover,to further refine the low-rank affinity matrix.We introduce the internal image statistical prior,called recurring pattern prior.This prior is originally used for image and video segmentation,and we extend it to the multi-shot Re-ID task to improve its robustness.