Research And Implementation Of Non-overlapping Multi-Camera Pedestrian Re-identification

Recently, with the development of intelligent video surveillance technology, person re-identification across non-overlapping camera views has drawn great interest in video surveillance. Person re-identification is to match pedestrian images observed in different camera views in non-overlapping multi-camera surveillance systems. Preson re-identication is a challenging problem, since a person's appearance often undergoes large variations across different camera views due to significant changes in low resolution, lighting, view angle, human pose and camera properties. To address these challenges, a novel method for person re-identification based on multi-feature subspace and kernel learning is proposed in this paper.The main work and innovations in this paper are as follows:1. Currently, distance metric learning-based person re-identification methods get the metric matrix in the original linear feature space directly, and then can obtian the function of the similarity between the samples. However, due to the linear original feature space can not be separated, the similarity and difference between samples can not be distinguished by the metric matrix obtained by the method mentioned above, which leads to poor recognition effect. This paper proposes an algorithm based on kernel learning, firstly, projecting the original feature space to a more easily distinguished nonlinear space through the corresponding kernel function, and then obtain the metric matrix M in the nonlinear space.2.The distance metric learning-based person re-identification methods often fuse multiple features by feature model and obtain the metric matrix through distance metric learning.However,the learned metric matrix often cannot sufficiently show the similarities and differences of samples, because this kind of method ignores the attribute differences between features. Therefore, this paper prosposes a method learning metric matrix individually, which can obtain the similarity function of different feature space.Furthermore, the similarity of samples are represented by combining different weighted similarity functions.3. Based on above innovations, a simple multi-camera person re-identification system is designed.