Research On Spatio-Temporal Feature Enhancement Video-Based Person Re-Identification

Pedestrian re-identification plays an important role in social public safety and smart cities,and is an indispensable part of achieving intelligent video surveillance.In practical scenarios,when the camera cannot get the face information and the resolution is not high,the pedestrian re-identification algorithm can continue to determine the identity of the target.It has important academic research value in the field of computer vision,and also shows important application value under the fields of intelligent security and intelligent people search in public places,unmanned shopping malls,human-computer interaction and intelligent photo albums,etc.Currently,image-based pedestrian re-identification algorithms have achieved encouraging performance.Video-based pedestrian re-identification,on the other hand,is attracting more and more researchers’ attention due to its rich spatio-temporal information.The presence of pedestrian occlusion,background object interference,and similarity of pedestrian appearance and pose in the pedestrian data collected in real scenes causes the model to fail to effectively extract better pedestrian appearance information using spatio-temporal information in pedestrian videos.In this thesis,the main work and research results are carried out for the video pedestrian re-identification spatio-temporal feature extraction problem as follows.(1)A video pedestrian re-identification method based on information bottleneck enhancement is proposed.Considering the data characteristics of video pedestrian rerecognition,the high content similarity of each frame of pedestrian images of video clips,and the existence of background interference information in each frame,a spatiotemporal feature optimization module(STIP)is designed to optimize the irrelevant information in spatio-temporal features.Information bottleneck loss is also introduced to maximize the mutual information between spatio-temporal features and identity labels of pedestrians,while minimizing the mutual information between features and inputs,reducing noise interference,and utilizing the effective temporal information in the video,thus improving the accuracy of pedestrian re-identification.Experimental analysis on the dataset MARS shows that the proposed method in this thesis achieves better recognition results and can effectively improve recognition performance.(2)A spatio-temporal feature matching attention video pedestrian re-identification method is proposed.Since the convolution operation focuses on the local information in time and space due to the size limitation of the kernel and cannot sense the information in the whole time and space,and the pedestrian images between frames are not perfectly aligned due to the pedestrian video in motion and the error of the pedestrian detector,the direct use of the 3D convolution operation to model the pedestrian appearance in the temporal dimension causes pedestrian appearance feature corruption.The feature matching attention module is introduced to enhance the acceptance of the global information of the model,while the local discriminative features of pedestrians are matched and attended to in the temporal dimension.A more general optimization framework for information bottleneck enhancement method is proposed for the same time on the information valve to further enhance the model.Experiments on the dataset MARS verify that the recognition accuracy and effectiveness of the proposed method in this chapter are improved.