Most current cross-modal pedestrian re-identification algorithms lack clustering ability and make it difficult to extract high-efficiency discriminative features; therefore, this paper proposes a multi-granular cross-modal pedestrian re-identification algorithm. First, the nonlocal attention mechanism module is added to the backbone network Resnet50 to focus on the relationship between long-distance pixels and retain detailed information. Second, a multi-branch network is used to extract fine-grained feature information to improve the distinguishing feature extraction ability of the model. Finally, the sample- and center-based triple losses are combined to supervise the training process, which achieves the purpose of accelerating the convergence of the model. The proposed method achieves Rank-1 and mean average precision of 62.83% and 58.10%, respectively, in the full search mode of the SYSU-MM01 dataset. In the visible-to-infrared mode of the RegDB dataset, Rank-1 and mAP reach 87.78% and 76.22%, respectively.