| Person re-identification（re-ID）aims to query or match the identical individuals across multiple non-overlapping surveillance cameras and has become an important visual task in various applications such as pedestrian security and video surveillance.For sake of changes in views,body poses,illumination,etc.,the captured images of the same person might vary significantly in different cameras.Thus,the re-ID is still a challenging visual task.Although much recent progress has been made to address the problem of the re-ID,the performance is insufficient and there is still much room for improvement.Henceforth,exploring the sound re-ID models would make great sense in both theory and practice.In the re-ID,the pedestrian images are captured by multi-view different cameras,thus it is reasonable to regard the re-ID task as a multi-view learning problem.In this regard,canonical correlation analysis（CCA）is a classical subspace learning method,which can be used to address this problem.Recently,various CCA-based variants have been proposed for the re-ID task.However,they have the limited performance and even are inferior to some traditional baseline methods.To this end,this thesis attempts to devise new discriminant terms and leverage meaningful prior knowledge,and further proposes CCA-based methods to compete to the state of the art.The main contributions of this paper are as follows:1.A ranking-embedded transfer CCA method（RTCCA）is proposed for person re-identification.Particularly,the triplet loss based ranking information is utilized to enhance the discriminant capability of CCA;meanwhile,the maximum mean discrepancy（MMD）as a regularization term serves as reducing the distribution discrepancy between the training and the test sets.These strategies make CCA more well-suited for the re-ID task.Experiments on VIPeR,PRID and 3DPES datasets are conducted,and have achieved the rank-1 accuracy to 50.3%,30.8%and 61.7%,respectively,which shows the effectiveness of the proposed model;2.A margin-embedding CCA with feature selection(l2,1-MCCA)for person re-identification is investigated.At first,we embed the margin information based on the contrast loss into CCA to enhance its discriminant capacity.Furthermore,we constraint CCA with l2,1-norm regularized terms over two projective matrices to identify the within-view correlative features for promoting the re-ID performance.To verify the efficacy of the proposed method,we also perform the re-ID task on VIPeR,PRID and3DPES datasets,and have achieved the rank-1 accuracy to 51.6%,29.7%and 62.5%,the experimental results show the effectiveness of our model. |
