Pedestrian re-identification is used to retrieve and recognize the same pedestrian in the non-overlapping fields of cross-camera. Aiming at the feature differences between different fields of cross-camera and the pseudo-label noise generated in the clustering stage, this paper proposes an attention-based unsupervised pedestrian re-identification method. In the feature extraction stage, an adaptive graph channel-spatial attention module (AGCBAM) is proposed, which considers both channel and spatial dimensions, adapts to cross-camera feature distribution by adaptively adjusting channel weights, and pays attention to specific spatial location features to capture details. In training stage, an improved intra-class neighbor spatial attention module is proposed, which combines label smoothing and spatial-level connections of positive instances to better remove pseudo-label noise and enable the model to better learn the real distribution of data. Through experiments on two mainstream datasets, Market-1501 and MSMT17, some existing common algorithms are compared, and the accuracy of mAP and Rank-1 is improved, which verifies the effectiveness of the proposed method.