Single-Shot Person Re-identification Based on Improved Random-Walk Pedestrian Segmentation Yu-Chen Chang Department of Computer Science National Tsing Hua University Hsinchu, Taiwan 30013 Email: fancla76@hotmail.com Chen-Kuo Chiang Department of Computer Science National Tsing Hua University Hsinchu, Taiwan 30013 Email: ckchiang@cs.nthu.edu.tw Shang-Hong Lai Department of Computer Science National Tsing Hua University Hsinchu, Taiwan 30013 Email: lai@cs.nthu.edu.tw Abstract—Single-shot person re-identification is to match pedestrian images captured from different cameras at differ- ent time under the condition of large illumination variations, different viewpoints, and inadequate information of single-shot case. To deal with these challenges, we propose a four-step single- shot person re-identification algorithm that consists of pedestrian segmentation, human region partitioning, feature extraction and human feature matching. Based on an improved Random Walks algorithm, human foreground is segmented by combining the shape prior information and the color seed constraint into the Random Walk formulation. Then color features of HSV histogram and 1-D RGB signal along with texture features from human body parts are used for the person re-identification. The correct match is then determined by the similarity scores of all features with appropriate weight selection. The experimental results demonstrate the superior performance by using the proposed algorithm compared to the previous representative methods. Keywords—single-shot person re-identification; Random Walks algorithm; pedestrian segmentation I. I NTRODUCTION There has been increasing interest on determining if a given person has been observed across disjoint camera views in surveillance system, known as person re-identification prob- lem. This aids for searching suspicious persons or robbers for police and guards in real life. The main challenge of this problem lays on pose variations, large illumination changes, different viewpoints, cluttered background, and low-resolution images. Previous work can be separated into two types which include single-shot case with only one probe and gallery image and multiple-shot case with a short sequence for each person. Due to the insufficient information, the single-shot case is harder than the multiple-shot case. In method [1], a one- against-all scheme was exploited to learn discriminatory model for single-shot person re-identification. This introduces the problem when a new person is given, it is not flexible for different gallery objects. Gray et al. [2] presented the method to find the best feature representation instead of designing novel features. However, a learning phase is required and the selected features may not be globally optimal because of sequential and independent selecting process. For multiple- shot case, Hamdoun et al. [3] proposed to match interest points to establish feature correspondences. Their work is simple and Figure 1. Flowchart of the proposed algorithms. fast, but it can not be used for dataset given pair images for each person. In method [4], an approach based on model fitting is presented to establish spatial correspondences and spatial- temporal segmentation for invariant signatures. This approach required consecutive frames as input data. It is, therefore, hard to extend for the single-shot problem. Because of these challenges and lacking of temporal infor- mation for single-shot case, we propose a two-stage algorithm with four steps, including human segmentation, human region partition, feature extraction and feature matching, as depicted in Figure 1. In our approach, two stages are introduced in the algorithm: (1) establishing correspondences and (2) searching correct feature matching for person re-identification. In contrast to the previous methods, the contribution of the proposed approach is twofold: First, a new method for single- shot person re-identification is presented without a learning phase. Therefore, it is more flexible and can be applied to large database. Second, based on a modified random walks algorithm incorporating constraints from shape prior and color distribution, a more accurate foreground mask can be obtained from pedestrian segmentation. This introduces a great benefit for the accuracy of person re-identification problem. Figure 2 depicts the results from the original Random Walks algorithm 1 2012 IEEE International Symposium on Intelligent Signal Processing and Communication Systems (ISPACS 2012) November 4-7, 2012 978-1-4673-5082-2 ©2012 IEEE