Abstract: High-quality 3D Gaussian human reconstruction from monocular video faces several challenges, such as pose discontinuity caused by skeletal articulation, geometric scale dependency in Gaussian optimization, and texture detail loss in Gaussian splatting. To address these issues, this paper proposes a temporal guid-ance and normal awareness method for high-quality 3D Gaussian human reconstruction. Firstly, based on full-body pose information from input video frames, the temporal-guided pose deformation transformed the canonical space initialization pose into an observation space pose through full-body articulated defor-mation. Then, a normal-aware Gaussian optimization was applied to densify and prune the Gaussian repre-sentation in the observation space, and the Gaussian representation of the human body can be achieved by optimized with normal maps supervision. Finally, shadow enhancement rendering combined shadow fea-tures extracted from normal maps with color information calculated by spherical harmonic functions, using Gaussian rasterization to produce the final rendered output of the Gaussian-represented human body. Ex-perimental results on the ZJU-MoCap dataset for novel view synthesis demonstrate that the proposed method achieves optimal performance in both average PSNR and LPIPS metrics, with totals of 189.95 dB and 171.6, respectively, outperforming comparison methods by 1.91 dB and 10.9, enabling high-quality reconstruction and real-time rendering of 3D Gaussian human bodies.