Person Re-Identification via Contextual Region-Based Metric Learning in Camera Sensor Networks

Abstract

Person re-identification in camera sensor networks is a challenging issue due to significant appearance variations of pedestrian images captured by different camera sensors. The contextual information of pedestrian images is a vital cue to overcome appearance variations. However, many existing approaches learn the distance metric in a global way or restrict to corresponding sub-regions, which discards the contextual information of pedestrians or learns the contextual information inadequately. In this paper, we propose an effective method to tackle the problem for person re-identification in camera sensor networks. Firstly, we propose the Contextual Region-based Metric Learning (CRML) to fully learn the contextual information in a local manner, which simultaneously utilizes three kinds of sub-region pairs to learn a discriminative transformation matrix. Secondly, we employ the greedy axis rotation algorithm to optimize the transformation matrix in the framework of mutual information. Thirdly, in the process of local similarity integration, we further propose the Context-Constrained Match (CCM) to overcome the misalignment problem by seeking the optimal match in the neighboring sub-regions. Fourthly, we further present the nCRML to avoid the dimensionality curse and fuse similarity scores in different low-dimensional subspaces. The experimental results on three challenging datasets (VIPeR, QMUL GRID and CUHK03) demonstrate the effectiveness of our method.