Multi-Label Joint Cross-Domain Person Re-Identification by Combining Post-Feature Extraction with Deep Mutual Learning Network

In recent years, deep learning-based person re-identification methods have significantly improved performance. However, environmental variations can lead to a substantial decrease in recognition accuracy, making unsupervised person re-identification necessary to address these challenges. Nonetheless, conventional person re-identification models fall short in realistic surveillance scenarios due to issues like occlusions in pedestrian images acquired by monitoring devices. This paper proposes a pose-guided unsupervised domain-adaptive recognition method to tackle the problem of pedestrian occlusion in unsupervised person re-identification. This approach effectively addresses the occlusion issues in person re-identification across different environmental conditions. Specifically, it utilizes a multi-label mutual learning cross-domain person re-identification, combining a pose feature guidance module with a deep mutual learning network. Furthermore, it introduces a high-response region keypoint filtering algorithm to eliminate noisy key points and guide the model toward improved feature representations. In terms of recognition accuracy measured by loss functions, it employs a multi-label training method to effectively enhance the model's results in each iteration and clustering process. The proposed model is compared with other state-of-the-art cross-domain person re-identification methods and achieves an accuracy of 70.2% in mAP metrics on Market-to-Duke unsupervised domain adaptation tasks. The effectiveness of its component modules is validated through ablation experiments.