Gait recognition is an identification technology based on walking patterns, with broad applications in fields such as video surveillance due to its non-contact and long-range recognition capabilities. However, real-world scenarios often pose challenges including viewpoint and clothing variations, and occlusion. Existing approaches predominantly focus on local feature extraction, often neglecting the differentiated dynamic characteristics exhibited by different body parts during motion, and fail to effectively model temporal dependencies across frames and handle occlusion interference. To address these issues, we present a Dual-Branch Adaptive Spatio-Temporal Aggregation Network (DASTANet) to learn more discriminative gait features. Specifically, the Spatio-Temporal Refinement Branch (STRB) captures global dynamic patterns and refines dynamic features by heterogeneous convolution. Complementarily, the Multi-Granularity Temporal-enhanced Branch (MGTB) incorporates adaptive temporal enhancement mechanisms with non-uniform chunking strategies to effectively extract local discriminative features. In addition, we design a Dynamic Region Mask (DRM) to enhance the robustness of the model in complex scenes by simulating multiple occlusion scenarios. Extensive experimental results on CASIA-B, OUMVLP and Gait3D public datasets show that our method outperforms other state-of-the-art methods. It achieves Rank-1 accuracies of 98.1%, 96.6%, and 90.4% on the CASIA-B dataset under NM, BG, and CL conditions, 93.0% on the OUMVLP dataset, and 56.4% on the Gait3D dataset.