Research And Implementationon Person Re-identification Method Based On Person High-order Feature

Person re-identification aims to identify persons with the same identity from images/videos captured by multiple independent cameras using computer vision methods.Person re-identification can distinguish persons with different identities from low-resolution environments by relying on the unique person features,which has a wide range of application scenarios in intelligent security,missing person search and other fields.In the actual scene,persons are often affected by the environment(e.g.the occlusion of surrounding objects,the change of light,and the different shooting angles),so how to effectively learn the appropriate features to distinguish the identity of persons is challenging.To this end,this study explores how to extract effective higher-order features for learning persons from different perspectives.(1)In Visible-Infrared Person Re-Identification,a large amount of information is missing due to the different modalities of person images.In order to make full use of higher-order person information,a method based on graph convolutional network and multi-hop attention mechanism is proposed.The person body is divided into several key-points,and the potential relationship between the key-points and the key-points is used to form the graph nodes in the graph convolutional network.The multi-hop attention mechanism is used to automatically strengthen/suppress some regional features to enhance the feature extraction ability of the network.The effectiveness of the method is verified on multiple datasets.(2)The Person Re-Identification from Aerial Perspective has the problems of excessive model parameters and poor recognition accuracy.A lightweight model based on multi-stage occlusion-generative distillation was proposed.A multi-stage occlusion-generation distillation method is constructed by using the occlusion-generation mechanism,which transfers the strong feature learning ability of the highweight network with large parameters to the lightweight model.And improves the higher-order feature learning ability of the model without increasing the parameter amount of the lightweight model.At the same time,the designed parallel network branch also has stronger higher-order learning ability than the traditional serial network structure.Experimental results show that the method has better recognition accuracy.(3)A Person Re-Identification analysis prototype system is designed,which is composed of front-end person acquisition module and back-end person re-Identification analysis module.The front-end person acquisition module obtains the data of the persons to be tested,and the system transmits the collected person information to the back-end analysis module for higher-order feature identification and displays the identified data.Then,a Person Re-identification dataset is constructed to verify the feasibility of the method in real scenarios.