Monocular single-frame 3D human pose estimation (HPE) has garnered significant interest, particularly in the domains of human-computer interaction and human action recognition. Several remarkable works have achieved excellent results in obtaining accurate 2D human poses. Building upon the foundation laid by previous advancements, this paper focuses on 2D-3D lifting. The proposed 2D-3D lifting algorithm in this paper is a Transformer-based model, which demonstrates its unique advantages in processing sequential data. The self-attention mechanism in the Transformer can handle global information without being limited by the receptive field. The Transformer's superior ability to process global information enables it to adaptively learn the relationships between human joints across different human behaviors. However, directly estimating 3D human pose from a single 2D image is a complex task due to depth ambiguity and joint occlusion. This ill-posed problem arises from the limited information available in a 2D image, making it challenging to determine the precise 3D pose. Additionally, occlusions further complicate the accurate estimation of joint positions. This article introduces a Transformer-based network that enhances the algorithm's robustness by utilizing multi-layer dual-stream blocks. One path concatenates the 2D coordinates and the focal length of the camera as input, while the other path takes the difference between the 2D coordinates as input. The two paths undergo an information fusion process at the front end of the block. Experiments conducted on various datasets verify the effectiveness of the algorithm and achieve state-of-the-art performance on Human3.6M and MPI-INF-3DHP benchmarks. Our code will be made publicly available on GitHub.