Research On Global Feature Enhancement Method For Person Re-Identification And System Implementation

With the increasing attention to the public safety,a large number of intelligent monitoring devices have been widely used in crowded places,such as squares and stations.As an indispensable part of intelligent monitoring systems,person re-identification has received widespread attention and quickly become a research hotspot.This paper focuses on the research of person re-identification based on global feature learning.Aiming at the current problems of global feature learning being susceptible to background noise interference and difficult to extract fine-grained global feature,a global feature enhancement network is designed,and a person re-identification system is constructed based on this network.The main research contents of this article are as follows:(1)Build a Point Attention ModuleAiming at the problem that global feature learning is susceptible to noise interference at present,this paper constructs a lightweight point attention module.This module can simultaneously model the global fine-grained correlation of feature map from both spatial and channel dimensions,thereby improving the anti noise interference ability of the network.In the spatial dimension,this paper designs a two-dimensional attention mechanism,which can continuously aggregate the feature information of the spatial dimension in the feature map and capture pixel-wise long-distance dependency in the spatial domain of feature map,thereby solving the problem of spatial dimension noise interference.In the channel dimension,this paper designs an external channel attention mechanism,which can model channel fine-grain correlation of feature map and adaptively calibrate channel feature responses,thereby solving the problem of noise interference in the channel dimension.(2)Design a Global Feature Enhancement NetworkThe point attention module can solve the problem of background noise interference,but it cannot take into account the extraction of fine-grained global feature in the process of extracting global feature.Therefore,this paper designs a global feature enhancement network based on the above methods.Firstly,the Inception-ResNet-v1 basic network is used to obtain strong multi-scale feature extraction capabilities which solves the problem of difficulty in extracting fine-grained global feature.Secondly,the Involution algorithm is used to capture long-distance interactive information about the spatial dimensions of the feature map,further improving the multi-scale feature extraction capability of the model.Thirdly,a Discrete Cosine Transform algorithm is used to replace Global Average Pooling to avoid significant loss of global feature information.Finally,embed the above point attention module to capture the global fine-grained long-distance dependency relationship of the feature map,it can improve the ability to resist background noise interference of the network and enhance the discrimination of global feature.(3)Design and implement a person re-identification systemThis paper designs and implements a person re-identification system based on the above research results.The system includes server,client,algorithm and astreaming media server.It implements real-time video monitoring,pedestrian detection,person re-identification and so on.It can meet the demand for retrieving target characters in images and videos.