Abstract: Existing millimeter-wave radar-based gait recognition methods are typically trained on small-scale datasets, resulting in poor model generalization and limited ability to extract effective global and local features from complex underground environments, which leads to low recognition accuracy. To address these issues, a gait recognition method for underground personnel based on the Denoising Diffusion Probabilistic Model (DDPM) and Multi-Branch Network (MBN) was proposed. The DDPM was used to denoise and augment the time–frequency spectrograms converted from radar echoes, effectively expanding the quantity of underground gait data and improving data quality. The MBN, consisting of one global branch and two local branches, extracted global gait features and local features of different granularities, enabling sufficient multi-scale feature extraction and improving the recognition of walking direction and speed. The Softmax loss and triplet loss were jointly employed to optimize coarse-grained features (2 048-dimensional features before dimensionality reduction) and fine-grained features (256-dimensional features after dimensionality reduction) in a collaborative manner, thereby enhancing the model's overall classification ability and feature discriminability. Experimental results showed that, on the self-built gait dataset, the DDPM-MBN model achieved Rank-1 accuracy and mean Average Precision (mAP) improvements of 8.05% and 16.96%, respectively, compared with ResNet-50. Compared with mainstream gait recognition models, the DDPM-MBN model achieved the best performance, with Rank-1 accuracy and mAP reaching 97.91% and 95.48%, respectively.