| With the rapid development of artificial intelligence technology,person re-identification has gradually become a popular research topic in the field of computer vision.As an important part of intelligent video surveillance system,it can help retrieve all the images from the gallery with the same identity as the query.Person re-identification technology can reduce lots of labor cost and has great research significance and application value.However,due to complex factors such as background perturbance,occlusion,low-quality images and different view angle,person re-identification faces severe challenges.In recent years,the mainstream image-based person re-identification methods mainly study how to extract more discriminative pedestrian features from the captured monitoring images,such as modeling the image feature relationship by attention mechanism,enhancing the recognition ability of local areas by pedestrian component matching,adding additional supervision signal auxiliary training,combining low-resolution and corresponding high-resolution images to enhance the cross-resolution re-identification performance,etc.However,there are still some problems,such as the single global modeling relationship,unified local division,depending on particular datasets,ignoring low-quality images and so on.How to extract discriminative foreground pedestrian features from pedestrian images under complex scene is still a challenging and far-reaching problem.This thesis takes the image-based person re-identification as the research object and do research on the application of attention mechanism,global features and local features in person re-identification under complex scene.In this thesis,three methods for person re-identification are proposed and implemented.The main research work and results are as follows:(1)This thesis proposes a non-local multiple-subspace attention-based method for person re-identification.In order to learn the second-order information and capture the various global relationships of the image,a Multi-Subspace Non-local Attention(MSNA)mechanism is proposed based on a multi-branch structure.MSNA is not limited to a single space and integrates local information into long-range dependency modeling by structuring the correlation of the different positions from different subspaces,so as to capture more comprehensive relationship information.Considering the global feature relationship,this thesis improves the ability of the model to enhance the features of foreground targets under complex scene and assign different weights to the features.Then,it gets the enhanced feature map with more prominent targets.The MSNA greatly improves the feature representation ability of the network in the low-level stage.The new models MSNA-Res50 and MSNA-MGN have achieved competitive results on popular datasets.In addition,the proposed MSNA is flexible and effective,and can be used as a general method to improve the feature representation of important areas in a plug and play manner.(2)This thesis proposes a reinforced loss based on random dropstripe for person re-identification.Based on the non-local multiple-subspace attention-based network,the person re-identification method is studied from the perspective of further strengthening local features.A Reinforced Loss(R-Loss)based on Random Dropstripe(RD)module is proposed to make full use of the information inside and across parts.It avoids the problem of ignoring cross-part information in uniform partitions,enhances the ability of local feature extraction,and pays attention to the correlation information between local features.The RD module adjusts the discarded or reserved areas in the feature map by setting different scale factors.Based on the scale-setting of RD,R-Loss takes different areas for person re-identification and forces the model to concentrate on the local detail features.When R-Loss is integrated into different levels of the model,the scale factor of RD is set differently to constrain the model to extract coarse-to-fine features.The new models R-MSNA-Res50 and R-MSNA-MGN combine the global and local information of pedestrians.Experiments on the popular datasets show that R-Loss can further improve the performance.In addition,the proposed R-Loss module provides a general method to enhance local features representation,which can be integrated into the network framework to improve the overall performance of the model.(3)This thesis proposes a cross-scale block-selection method for low-quality person re-identification.In the complex real-world scenarios,the research of person re-identification also faces the problems of low-quality images,such as low resolutions,blur and noise.In order to alleviate the above problems,a Cross-Scale Block-Selection Network(CSBSNet)with a single branch is proposed.Block-based attention is carefully designed,which involves neighbour pixels into computation and overcomes the interference of degradation factors.It mines the relationship between blocks.The proposed Cross-Scale Block-Selection(CSBS)module expands the block-based attention mechanism from single-scale to multi-scale,and explores the block-to-block correlation between cross-scale features.It can extract more "clean" and discriminative features.The model does not use low-quality and corresponding high-quality images,which reduces the dependence on specific datasets.The proposed method not only achieves competitive results on traditional datasets,but also verifies its robustness on a variety of low-quality images.Through the research on person re-identification under complex scene,this thesis proposes three effective methods.This thesis optimizes the network structure from the model level,enhance the anti-interference ability,and effectively extract the features of the target area in the complex pedestrian images.The proposed methods are compared and analyzed with the current mainstream person re-identification algorithms on several popular datasets,which verifies the great potential and superiority of the methods.In addition,they can be applied and popularized in more research topics in the future. |
