Person Re-Identification using Unsupervised Learning and ResNets in Surveillance Systems: A Progressive Clustering-based Learning Approach

Unsupervised person re-identification (re-ID) is a deep learning task aimed at extracting distinctive features without the need for labeled data, this task is crucial for practical applications in surveillance systems. Recent advancements in re-ID have shown the effectiveness of deep learning techniques in learning robust pedestrian representations. However, training deep models typically requires large-labeled datasets, which are both costly and impractical for real-world deployment. To overcome this challenge, we propose an Unsupervised Learning framework that transfers pre-trained deep representations to unseen domains with minimal reliance on labeled data. Our approach alternates between pseudo-labeling through image clustering and CNN fine-tuning, progressively refining feature representations. Additionally, we incorporate an advanced feature extraction strategy and k-means clustering for pseudo-labeling, ensuring robust unsupervised training. Extensive experiments on large-scale re-ID benchmarks, including Market-1501 and DukeMTMC-reID`, demonstrate that our method outperforms baseline models, achieving superior accuracy in both supervised and unsupervised settings. This framework offers a scalable and efficient solution for real-world person re-ID applications, effectively bridging the gap between supervised and unsupervised learning while significantly reducing annotation costs.