Dual-Cross Label Smoothing and Attention Driven Joint Multimodal Deep Learning Framework for Unsupervised Person Re-Identification

Abstract

The advancement of smart city infrastructure necessitates robust person Re-identification (Re-ID) systems capable of addressing challenges such as scalability, privacy, and security. This paper presents an unsupervised Re-ID framework that integrates enhanced data preprocessing, Efficient Net-B0 for feature extraction, K Means++ clustering for stable pseudo-labeling, a dual-branch discriminative learning structure, and Context-Aware Label Smoothing (CALS) to improve resilience to pseudo-label noise, occlusion, and viewpoint variation. The framework was evaluated on three complex datasets CASIA, Market-1501, and DukeMTMC-Re-ID each containing significant challenges such as pose variation, illumination changes, and background clutter. Experimental results demonstrate superior performance over conventional baselines including ResNet-50, DBSCAN, and Dual Cross-Neighbor Label Smoothing (DCLS). Both global and local learning branches achieved over 99% training accuracy within five epochs, indicating rapid convergence. The method achieved Rank-1 accuracies of 89.7%, 91.8%, and 87.5% and mAP scores of 82.5%, 85.7%, and 80.2% on CASIA, Market-1501, and DukeMTMC-Re-ID, respectively. Qualitative assessments and t-SNE visualizations confirmed improved retrieval accuracy and enhanced feature discrimination. The proposed approach demonstrates strong generalization, stability, and robustness against label noise, highlighting its suitability for real-world deployment in intelligent surveillance and public safety applications.