Person Re-identification In Surveillance And Forensics

Recently, more and more non-overlapping camera networks have been set up for monitoring pedestrian activities over a large public area, such as the airport, metro station and parking lot. In order to acquire individuals’ complete motion trajectories, matching persons across non-overlapping cameras in a surveillance camera network, a.k.a. person re-identification, is increasingly becoming a hot research spot in the computer vision community. Since traditional biometrics, such as face and gait, is unreliable or even infeasible in uncontrolled surveillance environment, body appearance is exploited for person re-identification in recent years. However, person re-identification remains an unsolved problem due to the challenges caused by view change, scale zooming and illumination variation, making different persons appear more alike than the same person in various cameras.In this dissertation, three key technologies in person re-identification framework are investigated, including feature representation, feature transformation and distance measure. Details are as follows:In feature representation, a sparsity based occlusion detection method is developed to address the problem that some patch can not matched at patch matching process when two person images are misalignment. Specially, two images captured by two d-ifferent surveillance cameras are generally misalignment, even though they are of the same person. Patch matching method, which matches the most similar patch pair, is used to address the misalignment problem. However, it ignore occlusion problem. In this paper, a occlusion detection is developed. Experiments results show that the proposed method can detect the occlusion patch effectively, and combination the occlu-sion detection methods with patch based matching methods improves the recognition performance about3-6%.In feature transformation, a feature projection matrix (FPM) method is proposed to directly project feature vectors from one camera to the feature space of the other camera to address the problem that the feature distributions of different cameras are not consistent. For learning the FPM, a supervised learning method is used, in which the objective function consists of two terms. The first acts to project images of the same person close to each other, while the second acts to take images of different persons a-part. With the proposed objective function, the FPM learning can be formulated as a smooth unconstrained convex optimization problem, where a simple batch gradient descent algorithm is utilized on randomly selected samples to efficiently solve the prob-lem without loss of accuracy. Extensive comparative experiment results have shown the promising prospect of the proposed method by directly compensating the device difference for person re-identification task, and the performance improvement is about10%.In distance measure, traditional distance functions are nor proper as they do not consider the distribution of samples. Due to introducing external sample supervised in-formation, the metric learning based method usually achieves better performance than traditional distance functions, and also be the study focus of this paper. However, existing metric learning methods generally use a uniform global distance function to test all images pairs, which may be inadequate to handle the complex real-world ap-plication where the individuality of data points needed to be specifically tackled. In this paper, a k nearest neighbor based distance function is proposed to measure the distance of two images pairs by their neighbors. The experiments results show that the proposed method performs comparatively or outperforms the-state-of-the-art person re-identification methods and the improvement is about3-5%.This dissertation addresses the practical application of surveillance camera and focus the person re-identification problem. The above threw technologies cover key steps for person re-identification. Even though these methods improve existing methods obviously, person re-identification is still very difficult in the actual application.