Person Re-identification Based On Fusion Attribute Feature

Person re-identification is an important computer vision task that aims to retrieve the target person from images monitored by multiple non-overlapping cameras.In recent years,with the improvement of national security awareness,the research of person re-identification has also been paid more and more attention by researchers.Person re-identification is often used in large-scale public occasions,such as suspect tracking,intelligent tracing,and unmanned supermarkets.As for previous research on person re-identification,researchers usually only focus on image-based methods.However,due to the increasing demand for person re-identification in real-world scenarios,image-based person re-identification research has been difficulty coping with the current situation.So,the researchers turn their attention to the video-based person re-identification research that is more in line with the actual scene.Video-based person re-identification can provide temporal information and coherent context information,which is more suitable for identifying surveillance camera videos with large amounts of data.First,image-based and video-based person re-identification research will face difficulties such as occlusion,illumination changes,perspective changes,and pose changes.These difficulties are needed to overcome urgently.Therefore,this thesis proposes to use person attributes as auxiliary information to assist person re-identification.To this end,this thesis designs a dual-branch network model including an attribute recognition branch and a person re-identification branch.In addition,different from the previous attribute-assisted person re-identification,this thesis proposes to divide the attributes into person attributes and scene attributes.This is because if the person attribute features are simply integrated into the person re-identification network alone,the effect of the entire person re-identification may be deteriorated,and the advantages of attribute features in person re-identification may not be fully utilized.In order to make the attribute recognition network suitable for video-based datasets,this thesis proposes to add a spatio-temporal module to the attribute recognition network,and sufficient experiments prove that this method can improve the performance of the attribute recognition network.Then,to further improve the superiority of attribute features to the entire model in attribute-assisted person re-identification,this thesis selects several attributes that are optimal for person re-identification according to many experimental results.At last,to pursue the fairness of the comparison with the current video-based person reidentification network,the baseline of the person re-identification branch in this thesis selects the currently commonly used network,and only makes a few changes.In order to further avoid the influence of the background in the pedestrian image on the overall recognition,this thesis proposes to add the strip pooling operation to the person re-identification network.The attribute-assisted person re-identification model proposed in this thesis can largely solve the difficulties mentioned above,and has high generalization and robustness.To verify the effectiveness of the above methods,this thesis conducts a large number of experiments on the two mainstream datasets,MARS and Duke MTMC-VID.The experimental results show that the method proposed in this thesis has strong competitiveness compared with the mainstream methods in recent years.