Unsupervised person re-identification is a computer vision method that identifies and matches pedestrians without any labeled data, utilizing feature extraction and clustering algorithms. To address common issues in current unsupervised person re-identification methods, such as insufficient feature extraction, inaccurate clustering, high computational complexity and lack of model robustness, this paper proposes a deep clustering learning-based approach for unsupervised person re-identification. First, we investigate the use of IBN-Net combined with Generalized Mean Pooling (GEM) as the feature extraction network, which enhances the discriminative power of the extracted features. Second, to mitigate the sensitivity of clustering algorithms to hyperparameters, we introduce the OPTICS algorithm to assist DBSCAN in selecting hyperparameters, thus reducing DBSCAN’s dependency on them. Additionally, to fully utilize all the data in the training set, outliers are treated as separate clusters and included in the initialization and updating process of the memory dictionary. Finally, to address the inconsistency in update rates among clusters in the memory dictionary, we propose a cluster-level memory dictionary that eliminates this issue. Experimental results validate the effectiveness of our approach, demonstrating significant improvements in both precision and accuracy in unsupervised person re-identification tasks.