To address the low accuracy of traditional stereo matching methods in depth-discontinuous and weak-texture regions, we propose an improved algorithm based on dynamic multi-feature fusion and dual-branch adaptive aggregation. A nonlinear weighting function dynamically integrates noise-resistant Re-Census, multi-directional gradient and Lab-color costs. Distinct arm extension rules are applied to weak texture and depth-discontinuous areas, enabling a dual-branch adaptive aggregation that adapts to local scene characteristics. Disparity estimation follows a Winner Takes All (WTA) strategy. The ultimate disparity map is generated through a region-based disparity optimization module and multiple optimization processes. Experimental results on the Middlebury dataset indicate a 21.7% relative decrease in the mismatch rate compared to the baseline. Evaluations on the KITTI dataset and real-world scenes captured with a ZED 2i camera further demonstrate the robustness of the proposed algorithm.