Research On Robustness Algorithm For Person Re-identification Under Pose Diversity

Person re-identification is an intelligent video image analysis technology for retrieving targets across multiple non-overlapping cameras.It has been widely used in video surveillance,pedestrian retrieval and other aspects to facilitate the construction of smart cities.However,the pose changes caused by the non-rigid structure of human body and the change of shooting angle make the person re-identification task more challenging.In order to achieve an efficient person re-identification algorithm,the paper conducts in-depth research on it and achieves the following achievements:(1)A person re-identification algorithm based on pose robust feature learning is proposed.In order to resist intra-class changes caused by pose variations,the algorithm obtains the final feature robust to pose variations by comparing and learning the pedestrian features under different poses.Firstly,the pose features extracted from other pedestrian images are used to generate new appearance features for the target pedestrian,so as to rich the appearance features of the target pedestrian under different poses.Then,these features are comparatively studied to obtain the final feature representation that is robust to pose variations.(2)A dual-stream network model combining pose diversity and local feature learning is proposed.Based on the first work,a method based on local feature learning is introduced.and a weight vector is designed to reduce the impact of pedestrian misalignment.The network model consists of two subnets: One subnet extracts features that are robust to pose variations;the other subnet captures rich pedestrian feature clues.Then,the fusion mechanism of the two subnet features is designed to further improve the performance of person re-identification.The first work mainly focuses on solving the problems caused by pose variations.Based on the first work,the second work uses the dual-stream network structure to fuse effective features to enhance the robustness of the network to pose variations.A large number of experimental data verify the advance of the work in this paper in solving the problem of pedestrian pose variations.