Metric Learning For Person Re-identification

Person re-identification refers to the retrieval of person under one camera whether the person appears in other cameras again in the non-overlapping multi-camera surveillance system.In the actual video monitoring scene,there is a great challenge to the person re-identification due to the influence of camera angle,illumination,occlusion and so on.In this paper,we further study the person re-identification based on metric learning.In order to solve the problem of sample linear inseparability in the practical application of metric learning,a method based on kernel metric learning is proposed.This method extends the linear metric learning method to nonlinear metric learning method by kernel method.In this paper,two kinds of person re-identification algorithms are proposed.The main contents are as follows: ? Aiming at the small sample problem and the real sample linear non separable problem in metric learning,the regularized linear discriminant analysis person re-identification algorithm based on chi square kernel(KRLDA)was proposed.The regularized linear discriminant analysis algorithm(RLDA)uses regularization to solve the singularity of intra-class scatter matrix,thereby reducing the impact of small sample problems.The KRLDA algorithm mapped linear inseparable input data into a high dimensional linear separable feature space using kernel function to obtain scatter matrix that describes adjacent data relationship.Then,regularized linear discriminant analysis was used to obtain low dimensional projection matrix for maintaining high dimensional separability characteristics to improve the recognition rate of the person re-identification algorithm.The experimental results show that the highest recognition rate of the algorithm is up to about 10% after the kernel mapping on the VIPeR dataset.The experimental results on other datasets have also increased by more than 5%,and the algorithm has a better effect compared with the classic person re-identification algorithm based on metric learning.? Taking into account the differences in the description ability of different characteristics,used the idea of multi-feature subspace to propose a kernel regularized linear discriminant analysis for person re-identification which is based on multi-feature subspace.Firstly,the algorithm maps the features which extracted from different feature spaces from the linearly inseparable original subspace to the linearly separable high-dimensional kernel space by using kernel functions.Then,two sub-projection matrices from high-dimensional to low-dimensional space are obtained by using regularized linear discriminant analysis.Finally,in the two projection feature subspace respectively calculated the distance between samples to get two distance measurement function,and used a linear combination weights to get the final distance measurement function.The experimental results on datasets verify the effectiveness of the algorithm and realize the influence of different weights on the performance of the algorithm.