In recent years, Transformer-based methods have shown excellent performance in the task of 3D human pose estimation. However, although existing methods can effectively model the long-range dependencies between joints through the global self-attention mechanism, they are prone to prediction biases of local motion trajectories in scenarios such as rapid limb movements, and there is a problem of insufficient modeling of local motion features. To address this issue, this paper proposes a Transformer architecture model that combines aconvolutional neural network (CNN) with a hybrid attention mechanism. By adding convolutional feature extraction, it significantly enhances the representation ability of local joint movements. A hybrid multi-hypothesis generation module (H-MHG) is designed to generate richer hypothesis information while taking efficiency into account, effectively making up for the deficiencies of traditional global perspective methods in capturing local dependencies. Subsequently, the self-hypothetical granular (SHG) module is used to further explore the diverse information in the data, ensuring that the model can capture more details. Finally, through the cross-hypothetical interaction (CHI) module, we fully integrate the feature information among different hypotheses, enhancing the robustness and accuracy of the model. Experimental results show that the performance of this model on the Human3.6M dataset is improved by 7.99% compared with the baseline model MHFormer, demonstrating the effectiveness of the proposed components and the overall architecture in the field of 3D human pose estimation.