Person Re-Identification In Non-Overlapping Camera Networks

Multi-camera tracking is a fundamental tendency for a distributed Modern video surveillance system, however the non-overlapping scenes between field of views interrupt the continuity of trails from different cameras. To overcome this problem, Person Re-identification technology is defined to establish correspondence between images and videos of a person taken from different cameras distributed over non-overlapping scenes.In this dissertation, we focus on Person Re-Identification in Non-Overlapping Camera Networks and propose three algorithms, including color feature construction cross specific camera pair, cross modality matching based on AdaBoost framework and Datum-Adaptive Local Metric learning for PRID. These methods can overcome difficulties of PRID, such as the changes in illumination, camera parameters, viewpoint and pedestrian posture. The main contribution of this dissertation can be summarized as follows(1) For a specific camera pair, the change of illumination condition and camera parameters causes the migration and distortion of color feature. To deal with this problem, we propose low rank matrix recovery for PRID. After analyzing the influence mechanism of illumination condition and camera parameters on color feature, we realize the linear relationship between color features of images from different cameras. Base on this linear relationship, we construct the color feature matrix with low rank. Moreover, from the probe image feature we can reconstruct its corresponding feature in gallery camera color space using low rank matrix completion. The reconstructed feature, which owns the same illumination condition and camera parameters with gallery images, are used to take place of original probe image for the nearest neighbor searching. This replacement eliminates the influence of the illumination condition and camera parameters change between cameras. Furthermore, we introduce a noise matrix to overcome the interference of illumination variations with time. Experimental results on VIPeR and CUHK02-P1 dataset demonstrate that our approach can deal with the influence of illumination condition and camera parameters change on color feature and improve the performance of PRID.(2) Under the influence of illumination, camera parameters, viewpoint, pedestrian posture and occlusions, the image feature between a specific camera pair lie in difference modalities and possess various data distribution differences. To copy with this situation, we propose cross modality projection based on AdaBoost framework. Using AdaBoost algorithm, we learn a set of cross-modality projection models, each of which consists of two projection functions, to deal with types of cross-camera data distribution differences and all the models complement with each other. We construct AdaBoost framework using two different projection methods separately, hashing projection function and linear projection function, and these two AdaBoost frameworks are fused together to get better performance. Experimental results on VIPeR and CUHK02-P1 dataset demonstrate the validity of our method.(3) Under the multi-camera network, due to significant changes in image configurations (i.e. combinations of view angle, lighting condition, pose, occlusion and etc.), each image has its own feature distribution. Facing such complex feature distribution condition, we propose a novel Datum-Adaptive Local Metric learning method for PRID, which learns individual local feature projection for each image sample according to the current data distribution and projects all samples into a common discriminative space for nearest neighbor search. We adopt an approximate strategy based on Local Coordinate Coding to learn local projections. Based on the local coordinate coding of each sample, its local projection can be approximated by the similar linear combination. Experimental results demonstrate that, compared with state-of the-art methods, the proposed approach obtains superior performance in dealing with complex feature distribution condition and model generalization.