Research And Implementation On Feature-Aware Person Re-Identification Method Based On Hybrid Attention

Person re-identification is a sub-problem of Image Retrieval.With the continuous updates of visual retrieval methods,person re-identification methods based on person images,video or verbal representations have started to be applied in urban security,intelligent people finding and other fields.However,due to the viewpoint difference and low resolution generated by different surveillance,as well as the problems of light change,occlusion and pose,the existing methods cannot accurately perceive person features;in addition,the existing cross-camera person reidentification system has lag in intelligent people hunting,so the research of person reidentification direction is still of great importance.Based on the above motives,this thesis proposes two hybrid attention-based person re-identification methods and designs a system for person re-identification based on realistic environment,the specific research work is as follows:(1)This thesis proposes a Multiple Feature Complementary Network(MFC-Net)to complementarily fuse redundant features.The MFC-Net retrieves pedestrians in the process of feature learning by Multiscale Channel Attention branch,feature computation and transformation by Weighted Positional Attention branch to obtain features with high attention on person location.Finally,the features of two attention branches are fused in the Multilayer Feature Fusion Module to effectively eliminate redundant effects.Extensive person re-identification experiments are conducted on UAV datasets and synthetic datasets,and the effectiveness of MFC-Net is verified.(2)This thesis proposes a Multi-aggregate Hybrid Attention Model(MHAM)to capture person detail features by focusing on spatial feature information through a Spatial Fusion Attention,combining information with different semantics obtained from different depths of the convolutional layer,and using crossover operations and aggregation operations for feature perception.The Branch Fusion Attention module aggregates the features obtained from the branch of spatial attention aggregation and the branch of contextual information fusion,and the effective branches are enhanced with attention to obtain more robust features.MHAM has been shown experimentally to improve feature representation and has shown leading results on several datasets.(3)This thesis designs a person re-identification system based on hybrid attention method is designed,which can find lost elderly people under different cameras in real scenes,draw a person walking road map according to the captured position of the lost elderly person’s image,and estimate the next walking direction according to the previous walking route,which can solve the lag of previous system tracking and provide great help to find lost elderly people,and the person re-identification system is also of great practical significance in catching suspects,etc.