Feature Guided Enhancement For Video-Based Person Re-Identification

With the rapid development of information technology,identity recognition has become an urgent demand in fields such as public security,access control,and security services.Meanwhile,with the widespread application of surveillance cameras,intelligent monitoring has become one of the important means to address social security challenges.Therefore,person re-identification has broad application prospects in practical applications.Image based person re-identification has achieved good results.However,these methods typically use single frame images containing limited information.In addition,image based methods are highly dependent on image quality,which poses significant limitations for their application in real-world scenarios.Unlike image based person re-identification,video-based person re-identification not only includes spatial information of characters,but also temporal information of characters.In recent years,with the emergence of large-scale video datasets,videobased person re-identification has received increasing attention.Although existing excellent methods have made effective progress,their performance severely deteriorates when faced with complex scenes,such as occlusion,and background interference factors.Secondly,most methods often only focus on the significant areas of pedestrians,resulting in local redundancy of video features and inability to distinguish pedestrians with similar appearances.Therefore,in response to the above difficulties,the main research work of this article is as follows:Firstly,to address the issues of pedestrian occlusion and local redundancy of video features,this thesis proposes a video-based person re-identification network based on feature guided enhancement.This method includes specific frame selection method,feature guided enhancement module,feature erasing branch,and global branch.The main goal of the feature guided enhancement module is to enhance the common features of pedestrians in video frames while preserving their own frame level features.Firstly,select a specific frame in the sequence using a specific frame selection method.Then,the spatial similarity between this frame feature and other frame features is calculated,and the weighted sum of attention and video frame features is used to reduce the influence of occlusion factors in video frames.Secondly,due to the fact that the global feature extraction branch usually focuses on significant but similar areas between different pedestrians,this will ignore the subtle but crucial differential information between different pedestrians.This thesis proposes a feature erasing branch that can drive the network to learn more comprehensive information by randomly erasing features in different regions of a video sequence.Secondly,in response to the issue of background interference,this thesis proposes a video-based person re-identification network based on pedestrian segmentation.This network includes a pedestrian segmentation module,feature erasing branch,and global branch.The pedestrian segmentation module segments the person in a specific frame,and then uses the mask to perform the weighted summation of the overall video sequence,so as to reduce the impact of background interference,and improve the network’s feature learning ability in combination with the dual branch structure.In order to demonstrate the effectiveness of the method,a large number of experiments were conducted on the two mainstream datasets of video-based person re-identification,Mars and Duke MTMC-Video Re ID.Among them,the algorithm proposed in this thesis achieved Rank1 accuracy of 89.3% and 89.5% on the Mars dataset,respectively.Rank1 accuracy reached 95.3% and 95.5% on the Duke MTMCVideo Re ID dataset.This indicates that the algorithm proposed in this thesis can effectively reduce the impact of occlusion factors,local redundancy of video features,and background interference factors on video-based person re-identification.Finally,this article constructs a corresponding video-based person reidentification system based on the above algorithm.