Research On Person Re-Identification Across Cameras

With the construction and improvement of large scale camera surveillance network, it has become increasingly important to identify and search a specific pedestrian over a camera network. Given a large candidate set of images or videos, person re-identification is to recognize some images or videos in that a specific pedestrian was involved. It is conducive to inquire his or her temporal and spatial cues for entrance or exit and establish semantic relationship among different scenes. The technology plays an important role in some applications like multi-camera object tracking, biometrics recognition, intelligent human-machine interaction, semantic analysis of surveillance videos, etc., and has been one of the most popular research topics in the field of multimedia signal processing and pattern recognition in recent years. However, it is a challenging problem to au-tomatically identify pedestrian images from different cameras. In particular, the complexity increases in proportion to the scale of the camera network. To solve these problems, in this thesis, we focus on investigation of the context knowledge exploitation of cameras, distance metric designing and learning, pedestrian image representation.Firstly, we propose a novel framework for person re-identification across cam-eras by using multi-task distance metric learning. In this framework, the context knowledge of cameras is exploited, one Mahalanobis distance metric is designed for each camera pair. On that basis, we propose a new method for person re-identification across cameras by learning multiple distance metrics. To alleviate overfitting, Inspired by multi-task learning in the field of machine learning, we formulate the multiple distance metrics learning as a multi-task distance metric learning problem. Furthermore, to validate the proposed framework, we propose a pairwise constraints based multi-task large margin distance metric learning model by modifying multi-task large margin nearest neighbor nearest neighbor algorithm to address a sparse set of pairwise similarity constraints in person re-identification datasets available. Experimental results demonstrate that the proposed framework for person re-identification across cameras by using multi-task distance metric learning shows a significant re-identification performance gain w.r.t. traditional person re-identification framework by learning a single distance metric.On the basis of the proposed framework using multi-task distance metric learning, regarding the specific characteristics of person re-identification as a Nearest Neighbors problem and a sparse set of pairwise similarity constraints in person re-identification datasets available, we propose a novel multi-task max-imally collapsing metric learning model and give its optimization approach by exploiting alternating optimization and Nesterov’s method. In addition to main-tain the original good properties of maximally collapsing metric learning model, the proposed model has two good theoretical properties that facilitate the op-timization of its objective function. For one thing, the objective function of multi-task maximally collapsing metric learning model is jointly convex w.r.t. multiple positive semi-definite matrices to be learnt. This property guarantees that the globally optimal solution can be found in theory. For another, the objective function has a Lipschitz-continuous gradient when an alternating opti-mization method is exploited to solve each sub-problem of the proposed model. This property guarantees that optimal first-order gradient can be used to solve each sub-problem. Experimental results demonstrate that our proposed multi-task maximally collapsing metric learning model works substantially better than other current state-of-the-art person re-identification methods, and the proposed multi-task maximally collapsing metric learning model achieves a significant re-identification performance gain w.r.t. pairwise constraints based multi-task large margin distance metric learning model proposed in the previous chapter.Finally, according application requirement for the context knowledge of cam-eras unavailable and multiple images from one pedestrian in the gallery set, we propose a generalized Earth Mover’s Distance for person re-identification. For one thing, we exploit the discriminative information of different regions from pedes-trian image in the gallery set, and propose discriminative Earth Mover’s Distance model for matching human body. We learn the discriminative models of all re-gions from each exemplar image through maximum margin criterion by exploiting the semantic discriminative information from exemplar images in the gallery set. For another, we consider the prior knowledge of the body configuration as the spa-tial constraints for matching pedestrians. Here the prior knowledge is defined as that the whole body is composed of upper body and lower one. On that basis, we propose a probabilistic map model to represent the configuration of human body. Bayes estimation method is exploited to adaptively pursue the probabilistic map for each pedestrian image. According to the pursued maps, the spatial incom-patibility is computed by Kullback-Leibler Divergence embedded into the ground distance of Earth Mover’s Distance. Experimental results demonstrate that the proposed generalized Earth Mover’s Distance substantially improve performance on person re-identification w.r.t. original Earth Mover’s Distance. Besides, to our best knowledge, the proposed generalized Earth Mover’s Distance works bet-ter on benchmark datasets (ETHZf, ETHZ2 and ETHZ3) than other current state-of-the-art person re-identification methods.