Research On Person Re-identification Using Metric Learning And Sparse Representation

Person re-identification, an important research branch in the field of intelligent videosurveillance, is developing rapidly but still needs to be improved in the current computervision field. It aims to identify all images of one pedestrian taken in different scenes fromdifferent cameras. Influenced by different views, scales and illuminations, there exists hugedifference among the features of pedestrian images, such as position, color, contour. Thus,how to improve the re-identification rate of pedestrian images is still a challenge. Accordingto the characteristics of pedestrian images, we propose a novel person re-identification modelusing metric learning and sparse representation. The main contributions of this model are asfollows.Firstly, this paper proposes a new texture feature extraction method – statistical denseSIFT(sd SIFT) for pedestrian image based on dense SIFT. Integrating the sd SIFT feature withHOG and the weighted joint color histogram from the images divided by stripes, a newpedestrian image descriptor is constructed. Combining color feature with texture feature, thedescriptor overcomes the vision influences caused by different views, scales andilluminations.Secondly, a novel pedestrian matching frame based on metric learning and sparserepresentation is proposed. In this frame, the feature distance ranking method contains severalsteps.(1) The Mahalanobis distance learning is introduced and the new image feature isobtained from the original descriptor, which is transformed by the positive semi-definitemetric matrix.(2) The new image feature is embed in the sparse model to evaluate thesimilarity between probe and gallery.(3) The rank of each person with non-zero sparsecoefficient in the dataset is obtained according to the standard reconstruction error, using theiterative re-weighted sparse ranking method. Using the Mahalanobis distance metric totransform the feature space of pedestrian, the accuracy of classification is enhanced. Besides,combination with iterative re-weighted sparse ranking guarantees the effectiveness ofclassification.Extensive experiments are designed and tested on the PRID 450 s and CAVIAR4 REIDdatasets. The experiment results on the PRID 450 s and CAVIAR4 REID datasets demonstratedby the performance metrics——CMC curve and rank-1, indicate that the proposed frameworkperforms better than state-of-the-art algorithms and improves rank-1 by about 5%, whichproves the effectiveness and robustness of the proposed algorithm.