Research On Person Re-Identification Method Based On Multi-Granularity Features

Person Re-identification(ReID)refers to the identification of the same pedestrian in cross-free surveillance videos at different times or places to achieve continuous tracking and re-identification of the target pedestrian.In recent years,with the development of computer vision technology and surveillance cameras,person re-identification technology has been widely used in public safety,intelligent transportation,retail,entertainment and other fields.However,due to the influence of misalignment of pedestrian images,complex background,blur,change of posture,change of perspective,change of image color,style and brightness,the effect of person re-identification is seriously affected.To solve these problems,this thesis studies the person re-identification method based on IBN-Net and channel attention and person re-identification method based on multi-granularity feature alignment and attention mechanism,and designs the corresponding person re-identification system.The research content of this thesis mainly includes the following parts:(1)In order to extract the appearance invariance features of pedestrians,IBNC-Net,a novel coarse-grained network based on IBN-Net and dual-pool channel attention module,is proposed in this thesis.Firstly,IBN-Net50-a is used as the backbone network to learn the features that do not change with the appearance changes of image style,color and brightness.Secondly,the dual pool channel attention module is embedded in different network layers to suppress irrelevant features and enhance discriminant features.Then,generalized average pooling is introduced and the pooling scale is automatically adjusted through model training.Finally,the proposed IBNC-Net method was verified on three popular data sets(Market1501,Duke MTMC-Re ID and CUHK03),and Rank-1 reached 95.6%,91.2% and 80.5%,respectively,and m AP reached 89.1%,80.3% and 79.4%,respectively.Experimental results show that the proposed method can effectively improve the accuracy of person re-identification model.(2)Aiming at person re-identification problems such as inaccurate detection frame and misalignment of local features in the scene of pedestrian posture changes,this thesis proposes a person re-identification method based on multi-granularity feature alignment and attention mechanism.Firstly,a spatial transformation network is introduced to align the input pedestrian images locally.Then,the model is divided into three branches,the global branch and two local branches.The global branch is combined with selective convolution kernel channel attention to learn the significant global semantics and background information of pedestrian coarse-grained discrimination.Finally,the local branch blocks the feature map horizontally to learn fine-grained features,and associates adjacent local blocks to solve the problem of semantic information loss and pedestrian feature misalignment at the edge of the block.Experimental results show that the proposed method can make full use of the multi-granularity features,effectively solve the problem of pedestrian feature misalignment and get good results on multiple data sets.(3)In view of the practical application requirements of person re-identification,this thesis designs and implements a person re-identification system based on deep learning.The system architecture is designed and the development of each module is realized.The pedestrian detection technology and pedestrian re-recognition technology are combined to realize the end-to-end application of pedestrian re-recognition.