The Research On Graph Embedding Matching Network Algorithm For Occluded Person Re-identification

Person re-identification(ReID),whose task goal is to recognize the same person image captured by different cameras,has become a research hotspot in the field of computer vision in recent years.However,in real life,the scene is complex and people are easily occluded by obstacles.Most the existing ReID methods focus on the overall image of the person,while ignoring the occluded image.When extracting features from the whole image,the extracted features may involve scattered information or partial features of occlusion.If the model does not distinguish between the occlusion area and the target area,it may lead to wrong retrieval results.Therefore,it is necessary to solve the problem of occluded person re-identification.In order to solve the more challenging problem of occluding person re-identification,first of all,this paper proposes the global discrepant contrastive pooling(GDCP)to obtain the global features including the discrimination of the target person image contrast information,and extracts the local features using the pose estimation(PE).On this basis,this paper proposes adaptive feature relationships(AFR)to combine local features to explore more differentiated local features and to supervise the learning of feature relationships.In the task of person reidentification,occlusion sometimes shows more discriminating information,resulting in mismatch of person images and reduced recognition efficiency.This paper proposes an graph embedded feature matching method(GEFM),which uses the graph convolutional network to achieve many-to-many soft alignment of feature information,extract more discriminating and robust complementary features,and improve recognition ability.In conclusion,this paper proposes a novel person re-identification model,which is named the Graph Embedded Matching Network(GEMN)model,and proves that the structure can better improve the recognition effect of person re-identification in the person re-identification dataset.The main work of this paper is as follows:(1)In this paper,a global discrepant contrastive pooling layer is proposed to supplement the traditional global max pooling and global average pooling technology,obtain the global features of the target person image containing contrast features information,remove noise,and weaken background interference.Experimental results show that the global discrepant contrastive pooling method is better than the traditional global max pool and the global average pool on the global feature extraction effect,and the person re-identification method has a significant accuracy improvement.(2)This paper proposes an adaptive feature relationship to extract more distinguishing feature information,solving the problem of no connection between feature information and single information.The relationship between various components is constructed,and the key points of the human body are used to determine the connection between structural supervision features,obtain the correlation between each feature information,and reduce the dispersion of image information.Learning the relationship between features promotes the information transfer of meaningful features of the non-occlusion area and inhibits the transfer of information of meaningless target person features due to occlusion or outliers.(3)In this paper,this paper proposes a relationship between each graph embedding node and node connection for feature matching method,avoiding using sensitive one-to-one hard alignment,rather than using many-to-many soft alignment,The non-occluded part of one image compensates for the occluded part of the other image,prediction of the similarity between the individual features of the two pair images.This feature information complementary approach effectively reduces the effects of feature misalignment and interference of different types of occlusion during the matching process,and improves the accuracy,robustness of local feature alignment.The graph embedding matching network proposed in this paper combines more discriminative global and local features.Features alignment achieved using graph matching network.Experiments show that our method is superior to existing person re-identification methods on Occluded_DukeMTMC,DukeMTMC-ReID and Market 1501 datasets.In particular.on the Occluded DukeMTMC dataset,Rank-1 achieved 60.5%and mAP achieved 49.2%.