Research On Cross-modal Person Re-identification Technology Based On Global And Local Feature Joint Learning

In recent years,the issue of public security has gradually attracted people’s attention,making the intelligent monitoring system plays an irreplaceable role in daily life.The pedestrian re-identification technology in the system has also attracted much attention.At present,the pedestrian re-identification technology is mainly studied from the identity matching between visible pedestrian images Due to the influence of camera and other factors,single-mode pedestrian re-recognition technology cannot match infrared pedestrian images.Therefore,how to eliminate the huge difference between visible and infrared images in cross-mode pedestrian re-recognition and extract discriminant pedestrian features is the main research problem of this paper.In order to solve the challenges of cross-modal re-recognition,this paper proposes a cross-modal pedestrian re-recognition method based on joint learning of global and local features to mine the similarity of pedestrian features in different modes In order to extract more rich low-level pedestrian characteristics under different mode,using the first two convolution block as the feature extraction of a particular mode,at the same time in order to eliminate the differences between different modal in the first two convolution block after embedding the modal differences ease module,the module through the guide channel attention the normalized to refine the characteristic figure,in keeping the figure identification ability of the same characteristics In order to enhance the expression ability of pedestrian features under different modes,this paper proposes a local mixed domain attention module,which is mainly composed of sequential cascade of channel attention and spatial attention,and focuses on extracting discriminant local features In order to learn the common features of cross-modal pedestrian image re-recognition under different modes,this paper adopts global graph attention to train the model.The main purpose is to extract the global shared features across modes by clustering the features of adjacent nodes under two modes,thus improving the recognition rate of cross-modal pedestrian re-recognition.This article uses ID losses The triplet loss and center loss were optimized on data sets SYSU-MM01 and RegDB,and the whole model was trained using dynamic aggregation learning strategy.Experimental results show that the proposed method not only eliminates the modal differences,but also significantly enhances the pedestrian feature expression ability and improves the identification accuracy of target pedestrians.