Research On Robust Person Re-Identification Algorithm Based On Graph Representation

Person Re-Identification is a hot research direction in the field of computer vision.It mainly solves the recognition and retrieval of pedestrians across cameras and scenes.It has important research significance and broad application prospects.This technology can also be used as an important supplement to face recognition technology,for cross-camera continuous tracking of pedestrians who cannot obtain a clear face.In this thesis,the graph representation of human joint posture is used as auxiliary features,combined with deep learning and graph reasoning,and the problems of insufficient feature extraction of pedestrian re recognition,low accuracy of small-scale pedestrian recognition and the neglect of the relationship between key points are studied and solved.Three network models are proposed contrapuntally.The research results of this thesis are as follows:(1)A Multi-Scale Feature Fusion Network for Re ID(MS-FFN)is proposed.MSFFN integrates features of different scales to improve the pedestrian re-recognition performance of the model,enhance the representation ability of features,and ultimately improve the performance of the model.The experimental results on the Occluded-Duke dataset show that compared with the current state-of-the-art algorithms(such as HONet),the accuracy of R1 of the proposed MS-FFN model is increased by 0.5%.(2)A Locational Attention Network for Re ID(LANet)is proposed.LAnet embeds location information into channel attention and jointly learns location and channel features,thereby improving the performance of pedestrian re-identification algorithm models.The experimental results on the data set Occluded-Duke show that compared with the current state-of-the-art algorithms(such as HONet),the R1 accuracy of the proposed LANet model is improved by 1.1%.(3)A Graph Reasoning Network for Re ID(GRNet)is proposed.GRNet maps the features to the interactive space,then uses graph convolution for relational reasoning,and finally maps the features with relational information back to the original coordinate space to obtain a deeper relational representation.The experimental results on the data set Occluded-Duke show that the proposed GRNet model is better than the previous LANet,MS-FFN and HONet models,and the R1 accuracy of the model is increased by 3.1%,3.7% and 4.2% respectively;The results on the data set Market-1501 show that compared with state-of-the-art algorithms(such as PGFA),m AP and R1 are increased by 5.9% and1.5%,respectively.