Research On Person Re-identification Based On Metric Learning

Person re-identification refers to the research of identifying and matching the same person from different camera angles.This technology is mainly used in intelligent video surveillance and has important academic significance and application value in the field of computer vision.Due to the rapid development of big data in recent years and the urgent need for intelligent video analysis technology,this paper focuses on the similarity measurement technology in person re-identification starting from metric learning.From the perspective of single mode and cross mode,two different methods are innovatively proposed to effectively improve the accuracy of person re-identification.Starting from the learning direction based on the joint enhanced local maximum occurrence feature and k-KISSME metric,many challenges in the current person re-identification,such as similar appearance of person,changes in misaligned viewing angles,etc.,and the same person under different characteristics feature extraction leads to the metric difference and the problem of inaccurate matching.This paper proposes a person re-identification method that jointly enhances the local maximum occurrence feature and k-KISSME metric learning.In the feature representation stage,jointly enhancing the local maximum occurrence feature can simultaneously obtain the pedestrian's global appearance feature and local detail feature.Effective feature representation is robust against changes in light and viewing angles,and discriminative metric learning can improve the matching accuracy of person images.Enhanced effective representation of local maximum occurrence can achieve the fusion of fine details of person images and overall appearance information to meet the human visual recognition mechanism;the KISSME metric learning(k-KISSME)method is simple and efficient,and only requires estimate the two inverses covariance matrix.In addition,in order to deal with light and viewing angle changes,we use Retinex transform and scale-invariant texture descriptors.The experimental results after comparison with the related methods in three widely used datasets show that the algorithm proposed in this paper not only has a rich ability to express person features,but also improves the recognition rate of person re-identification.From the perspective of cross-modal person re-identification,how to build a model that can extract person appearance feature representations in two different modalities and how to effectively use person appearance features in two different modalities to calculate.The degree of similarity between person is two important issues.This paper designs a cross-modal person re-identification architecture based on visible light images and infrared images.First,the paper constructs an end-to-end dual-path multi-branch cross-modal network that uses the inherent relationship between RGB and infrared images to transfer the embedded representation from one mode to another.Then,the paper introduces the Multiple Granularity Network(MGN)architecture to obtain distinctive cross-modal feature representations,and improves the expression ability of feature embedding by combining local and global information of the image.Experimental results on the latest SYSU-MM01 dataset show that the method has better performance than the existing mainstream methods on the RGB-IR benchmark person re-identification dataset.