Research on Unsupervised Pedestrian Re identification Algorithm Based on Domain Adaptive Method of IBN Network and Label Denoising

Abstract

This study focuses on the problem of Person Re-Identification (ReID) and deeply explores the application of unsupervised domain adaptation methods in this field. Person re-identification technology is crucial for retrieving specific pedestrian images across cameras and is a core technology for intelligent video surveillance and smart security. Although deep learning has driven significant progress in person re-identification, current methods still face challenges such as scarcity of data annotation and domain differences. To address these issues, we propose a combination of Instance Normalization and Batch Normalization, known as the IBN network module, and a method that utilizes clustering to generate pseudo-labels, thereby implementing person re-identification within an unsupervised domain adaptation framework. We review related technologies and conduct an in-depth study of datasets and evaluation metrics for unsupervised person re-identification. In the experimental section, we validate our approach using the Market-1501 and DukeMTMC-ReID datasets. The results show that both the IBN module and the unsupervised domain adaptation method can effectively improve performance. Additionally, we propose an unsupervised person re-identification method based on label denoising to further enhance accuracy. Although this study has achieved certain results, real-time person re-identification, efficient and robust network design, and the application of unsupervised learning remain focal points for future research.