Research On Person Re-identification Algorithm Based On Metric Learning

Person re-identification refers to the problem that how to find the same person in the existing person image database for a given person image,usually these person images are derived from non-overlapping view of the camera.This problem is a key task in intelligent surveillance,which can be applied to intelligent security and provide clues for searching pedestrians.It has very important practical significance and becomes the research hotspot in the field of computer vision.Because of the change of views,change of illumination,different posture,the image of the same person taken by different cameras may be different in apparent characteristics.How to improve the accuracy of recognizing person image is a big problem.In this paper,the quality of the person image which checked from the person re-identification database is low,and the feature of the person between the same pedestrians may be changed.We use the MSRCR-G algorithm to enhance the image quality which is improved on the MSRCR algorithm.The algorithm is based on the color constancy theory,which can reduce the influence of changes in light.At the same time,it should ensure the person image's color is not distorted after be processed.In order to solve the complex matching problem in person re-identification,this paper proposes a cross-view secondary discriminant measurement learning algorithm.First,enhance the person images,then extract the LOMO feature.We use the ross view regular discriminant analysis method to study a sample subspace,At the same time,a suitable distance metric function is learned in the person character subspace.The method can maximize the inter-class distance and minimize the intra-class distance of person images.At the same time,on the basis of the original algorithm,the singularity problem of covariance matrix in the process is solved effectively.In this paper,the two algorithms are experimented separately.We use the evaluation standard such as the average brightness,information entropy,contrast,histogram.to measure the quality of the images which are processed by the MSRCR-G algorithm.The person images before and after be processed by MSRCR-G algorithm are compared from the data and visually.The results are gained by experimenting from two challenging person re-identification datasets which called VIPeR and GRID,shows that the proposed XRDA distance measurement algorithm improves the recognition rate of person re-identification.Experiments show that this algorithm can effectively reduce the influence of light changes on person appearance which also solve the problem of color distortion.The accuracy of person recognition is improved.Finally,we discussed the advantages and disadvantages of this algorithm and the future work direction.