Person Re-identification Based On Coupled Feature Spaces Learning

Because of the high attention to the public security, surveillance systems have been used more frequently. As the public space camera networks have grown quickly in recent years, it is becoming increasingly clear that manual re-identification is prohibitively costly and inaccurate. There is a growing interest in developing automated re-identification solutions in the field of computer vision community. Person re-identification is to match pedestrians observed from non-overlapping camera views based on image appearance. This is a very challenging problem because images of the same individual can be very different due to variations in pose, viewpoint, and illumination. Solving the re-identification problem has gained a rapid increase in attention in both academic research communities and industrial laboratories in recent years.Two fundamental problems are critical for person re-identification, feature representation and metric learning. To address the limitations of featuring different static and dynamic backgrounds, the main contributions of our thesis can be summarized as follows:We proposed an efficient metric learning method called joint graph regularized coupled feature spaces learning to improve re-identification accuracy. Metric learning is widely used to model the transformation among cameras recently. Our method learns two projection matrices to map multimodal data into a common feature space, in which picture of the same pedestrian from different cameras can be performed. And in the learning procedure, the l2,1, norm penalties are imposed on the two projection matrices separately, which leads to select relevant and discriminative features from coupled feature spaces simultaneously. A trace norm is further imposed on the projected data as a low-rank constraint, which enhances the relevance of different modal data with connections. We found that the locality constraints in projection metric are helpful for metric learning, so we try to integrate the structure of the target images into a joint graph regularization to exploit the structure information. The projection metrics are complementary to each other in the joint graph regularization and optimizing the images of the same person from different cameras simultaneously.Comparisons results show the superiority and efficiency of our proposed method with performance measured in terms of Cumulative Match Characteristic curves (CMC) on two challenging datasets. We applies the method of pedestrian re-identification in this thesis to a re-identification system and tracking technology for counting and risk, the system can find out the corresponding images of the pedestrian who is the target by importing in the candidate set.