Research On Person Re-Identification Based On Attention Model And Feature Level Affine Alignment Model

Camera is everywhere in modern cities,which produce huge amounts of surveillance video data every day.In surveillance video,human analysis is the most important.Person Re-identification(Re-ID)is one of the important research issues,and it is also one of the hotspots in the field of computer vision in recent years.It aims to match people across non-overlapping camera views in an image or video.Its research has broad application prospects,including criminal investigation in the field of hunting criminals,searching for missing persons and pedestrian statistics in malls.The prosperity of deep convolutional network has introduced more powerful representations with better discrimination and robustness for pedestrian images,which pushed the performance of Re-ID to a new level.But there is still a certain distance to achieve the goal of practical application.The above problems need further research to solve.In this paper,firstly,we summarizes the research methods and status of Re-ID at home and abroad,including from traditional manual design feature model and metric learning methods to current mainstream methods based on deep learning,and analyses the main challenges faced by Re-ID tasks.Secondly,we proposed two novel models based on existing approaches,and the feasibility and effectiveness of the algorithm are demonstrated on some challenging open datasets,Market-1501,Duke MTMC-reID and MST17.The specific work of this paper is as follows:For pedestrian with large pose variations,complex background clutters,severe occlusions,we proposed person re-identification based on regional attention.Because ResNet-50 has better pedestrian localization characteristics,regional attention is used to further learn pedestrian saliency semantics features and obtain pedestrian features with stronger robustness and higher discrimination against gesture changes.The model structure is simple and easy to implement without requiring additional supervision information.Finally,the experimental results fully demonstrate the effectiveness and robustness of the proposed model.Due to unreliable detection lead to misalignment,refined person re-identification model based on affine alignment of feature layer were proposed.Using convolutional networks to learn affine parameters,automatically align pedestrians,and then learn well-aligned pedestrian local features.Such local features are more suitable for pedestrian similarity matching,which solves the limitations of using global features and reduces misalignment problem significantly.Meanwhile,improves the recognition accuracy.