In the field of computer vision, the task of unsupervised person re-identification presents significant challenges. The features extracted by mainstream unsupervised methods are greatly affected by background noise and lack effective strategies to guide clustering optimization, which is not conducive to learning high-quality pseudo-labels. In order to mitigate these challenges, this paper introduces a Pixel-level Fine-grained Feature Attention (PFFA) that extracts pixel-level fine-grained features of person characteristics in both horizontal and vertical directions in space, and highlights person features by merging fine-grained features in both directions, thereby reducing the impact of background noise. Secondly, during the clustering optimization phase, we present a Hard Contrastive Learning (HCL) to facilitate model learning by selecting hard positive samples to form contrastive loss for clustering optimization, enhancing the correlation between intra-class samples by converging these samples towards positive proxies. Finally, We carry out comprehensive experiments and in-depth analysis on the model. The findings demonstrate that the newly proposed method attains 80.4% mAP and 92.4% R-1 accuracy on the Market-1501 dataset; 80.4% mAP and 92.4% R-1 accuracy on DukeMTMC-reID; and 69.3% mAP and 83.0% R-1 accuracy on MSMT17 dataset.