Person Re-identification Based On Visual Salient Feature Analysis

The problem of person re-identification mainly refers to the matching among wide area camera networks, especially in non-overlapping camera networks. It is one of the important tasks of social security and intelligent surveillance. However, person re-identification in camera networks is a significantly challenging task since the same pedestrian in different camera views often undergoes variations of postures, appearance, viewpoints, and illumination.The primary process of person re-identification includes the extraction and fusion of human images or videos, and the design of matching model to compute the similarity score. According to the input of measurement model, the researches on person re-identification can be divided into the single-shot case and the multi-shot case. The measurement model of single-shot case each person only has one image in a camera view as input data. On the contrary, multi-shot case has multi-images or video as input data.This thesis focuses on research on person re-identification in two cases: single-shot based and multi-shot based scenarios. Firstly, towards the person re-identification in the single-shot case, we regard the salience information of person as visual salient feature of human images, while salience learning is usually understood as to find the regions which are salient of original image. As salience information is important for person re-identification, how to find visual salient regions of human images is crucial to salience based methods. Furthermore, towards the person re-identification in the multi-shot case, we regard these key frames as the visual salient feature of human video sequences. Existing methods which based on the matching among single images are often ignored information contained in person image sequence, and in practical the data of camera networks are stored in video, that is multi-shot. How to discriminative and key image sequence from original video is key point towards these video-based methods. Aiming at these problems, the main contributions and innovations of this thesis are as follows,Firstly, towards the problem of presentation of visual salient feature of person re-identification in the single-shot case, we propose a novel model for salience learning, and this method is based on outlier detection and by using distance and density information to compute image salience value. In this work, we firstly defined the salience learning as one kind problem of outlier detection, and this method combined the density and distance information of local dense patches for salience learning.Secondly, on the basis of getting the visual salient feature of human images, we give a new model of person re-identification in the single-shot case. The model combined salient regions and prior estimation of human head location to compute similarity scores between patches pairs. The effectiveness of our model is validated on some benchmark datasets, which demonstrates successful results and efficiency.Thirdly, as the visual salient feature of person re-identification in the multi-shot case, we design the framework by applying simultaneous sparse recovery method for representative/key frames extraction and these key frames are discriminative images in video. The discriminative images selection is achieved through the solving row-sparsity regularized trace minimization problem. Compared with single human image, multi-key frames contain richer information which can remove the inference of occlusion and generate better performance.