Abstract: To address the challenge of balancing occlusion robustness and pixel alignment accuracy in 3D human pose and shape estimation, a dual-branch estimation method integrating attention mechanisms is proposed. The proposed method consists of a segmentation-guided branch and a pose construction branch. The segmentation-guided branch adopts a human body part segmentation strategy, generating segmentation results for different body parts and converting them into attention weights to guide feature learning. The pose construction branch performs human model parameter regression based on the input feature maps combined with the attention weights, achieving accurate 3D human pose and shape estimation. Compared with state-of-the-art methods, the proposed method achieves improvements on the Human3.6M dataset, reducing MPJPE and PA-MPJPE by 5.4 mm and 1.9 mm, respectively, and on the 3DPW dataset, reducing MPJPE and PVE by 0.2 mm and 3.2 mm, respectively. Experimental results demonstrate that the proposed method not only significantly improves the alignment accuracy between the reconstructed mesh and the input image but also exhibits enhanced occlusion robustness in complex scenes.