Research On Person Re-identification Method Based On Multi-level Feature Fusion

With the continuous development and improvement of video surveillance networks,the amount of data in the field of security monitoring is growing rapidly,which makes the traditional method of cross-regional pedestrian tracking through manual comparison no longer sufficient.Person re-identification(Re-ID)technology aims to achieve accurate identification and matching of the same pedestrian in different surveillance videos.By associating the target pedestrian’s image with the collected pedestrian images from cameras,person Re-ID technology can replace manual completion of cross-device retrieval tasks,having a broad application prospect in the fields of intelligent security and others.However,in actual scenarios,changes in factors such as camera shooting angle,shooting environment,and pedestrian posture and clothing can affect the accuracy of person Re-ID models.Improving the richness and diversity of pedestrian features extracted by the model is the key to improving accuracy.In this paper,based on multi-level features,using feature fusion,multi-granularity block,attention mechanism and Trans Former,the feature extraction ability of the model is enhanced,the accuracy of the model is improved,so as to adapt to complex and changing person Re-ID tasks.The research content of this paper is as follows:(1)This paper proposes a person re-identification method based on Progressive Attention Pyramid(PAP)to fuse multi-level features.To address the issue of interference during the fusion of multi-level features,this paper proposes a PAP structure,which is designed to fully extract effective features from multi-level feature maps and suppress irrelevant interference.The channel attention module is embedded into the PAP to form the channel attention pyramid module,which is applied to the channel dimension of feature maps at different levels of the backbone network to aggregate multi-level key channel features.The spatial attention module is embedded into the PAP to form the spatial attention pyramid module,which is applied to the spatial dimension of multi-level feature maps to extract multi-level and multi-granularity pedestrian detail features.The PAP adopts a multi-level structure,and each level captures effective information from fine-grained to coarse-grained based on the "split-attend-merge" principle in the multi-level feature maps,enabling the model to obtain a comprehensive,complete,and integrated perception of pedestrians.(2)This paper proposes a person re-identification method based on multi-level feature fusion and multi-graph interaction modeling(MLFF-MGI).When fusing the output feature maps of various levels in the backbone network,the simplistic fusion approach may lead to the loss of pedestrian detail information in shallow feature maps.To address this issue,this paper proposes a bi-directional residual connection pyramid structure for fusing multi-level features,and uses a progressive attention pyramid module and cross-orthogonal regularization to enhance the shallow detail features.The feature alignment module is used to guide the fusion of shallow feature maps and mitigate the impact of feature misalignment on model accuracy.In addition,this paper introduces the Transformer to model the interaction relationships between multiple pedestrian images,in order to extract more stable common features of pedestrians,suppress the interference of exceptional features on the model,and further improve the accuracy of the model.(3)Based on the proposed person Re-ID algorithm in this paper,a person tracking system is designed and implemented,providing a useful reference for further promoting the application of person Re-ID technology.The system consists of three core algorithms: person Re-ID,person detection,and person attribute recognition,which are implemented using technologies such as React,Flask,and My SQL.Users can upload images of target persons or set their attributes for surveillance.The system retrieves target persons based on real-time data from monitoring videos or cameras.After the target person is retrieved,the system reconstructs their trajectory and saves log information for user inquiry.Additionally,the person tracking system includes a person gallery page that summarizes all captured person images.Users can filter the images by person attributes and support one-click surveillance deployment.