| Person Re-Identification(Re ID)is one of the popular research directions in the field of computer vision,and has broad practical application prospects in scenarios such as smart life,financial analysis,and public safety.In recent years,researchers have achieved great success in the task of pedestrian re-identification,and the research method based on Convolutional Neural Networks(CNN)is particularly attractive.However,with the rapid development of computer Transformer technology in the field of computer vision,many researchers began to explore how to apply it to the task of person re-identification.These researchers proposed a series of excellent model structures,and further adapted and adjusted them to improve their performance in person re-identification tasks,and achieved excellent retrieval performance.According to different scenarios,person re-identification tasks can be divided into two types: standard environment and occluded environment,and key problems need to be solved respectively.The key issue in the standard person re-identification task is how to extract robust and semantically-rich pedestrian image features.The key issue of the occluded person re-identification task is how to make the model better learn the ability to distinguish human body parts and occluded parts,so as to detect unoccluded pedestrian areas,and then perform feature matching with unoccluded full-body image images.Based on Transformer technology,this thesis proposes corresponding improvement methods for two pedestrian re-identification tasks.The main research work is as follows:1.For the standard person re-identification task,a person re-identification method based on the Transformer structure is proposed,which aims to fuse multi-scale features to improve model performance.Input the image into the CNN network,and use Transformer to extract global information from multiple scale blocks in the CNN network structure.In the process of feature transfer at each scale,a residual structure is added to retain certain features with local information,and feature splicing is adopted to retain features with global information at the upper level to reduce the information in the process of feature transfer Loss,so that the final pedestrian features more completely retain the semantic information in the low-level scale blocks.This method makes full use of the advantages of the two network structures of CNN and Transformer,and effectively improves the ability of the model to fuse multi-scale features.2.For the occluded person re-identification task,an occluded person re-identification method combined with pose estimation is proposed.This method obtains the pedestrian heat map through the pose estimation model,and combines the Transformer model to obtain local features with global discriminative power.Through local feature matching and measurement,the distance between human body part features and occluded part features is shortened,the interference of noise information caused by obstacle occlusion is reduced,and the ability of the network to pay attention to pedestrian information is improved.The research finally demonstrates the effectiveness of the proposed method by conducting extensive experiments on occluded person re-ID datasets and standard person re-ID datasets. |
