Abstract: To address the problem that the clustering results of unsupervised person re-identification methods contain a large amount of noise in the pseudo labels, a person re-identification method combining reliable instance mining and features optimization is proposed. Firstly, an indicator is designed to measure the reliability of pseudo labels by using the stability of DBSCAN clustering results under different parameters. Secondly, a reliable instance mining strategy is proposed to denoise the pseudo labels. Instances with pseudo label credibility greater than the preset threshold retain their original pseudo labels, otherwise, corrected their pseudo labels. Thirdly, a dual momentum update strategy is adopted to update the global and local features, i.e., each batch updates the features of the involved samples instantly, and each epoch updates the features of all samples in the memory bank.Finally, unified contrast loss is used to train and optimize the backbone neural network. Experimental results on two large public datasets, Market-1501 and DukeMTMC-reID, showed that mAP reached 77.9% and 67.4% respectively, and Rank-1 reached 90.2% and 88.2% respectively.