Accurate 3D human pose reconstruction from a single RGB image remains challenging due to scale ambiguity and perspective distortions. Current single-view methods primarily rely on learned priors or kinematic constraints, but they often struggle to maintain geometric consistency with the physical scene. This results in horizon alignment drift and instability when rendered in metric environments. To overcome these limitations, this study introduces a vanishing-point-driven framework that integrates scene geometry into the pose correction process. Under the Manhattan-world assumption, dominant vanishing points are detected to estimate the ground plane and recover the camera orientation with high precision. A lightweight 3D pose estimation network generates initial joint coordinates in camera-centric space. These coordinates are then refined through a VP-based ground-alignment transformation, which resolves scale ambiguity and minimizes geometric drift. The corrected poses are normalized to physical scale and streamed to NVIDIA Omniverse^TM for real-time digital-twin visualization. Experiments conducted on indoor scenes from the NYU Depth V2 dataset demonstrate sub-pixel accuracy in vanishing-point localization and significant improvements in geometric alignment between the reconstructed poses and the true scene layout. This confirms the effectiveness of the proposed approach for single-view digital-twin human modeling.

In this study we investigated Kalman filter-based attitude estimation algorithms, considering external acceleration and bias effects in several different scenarios. Towards these goals, gyro biases were first estimated, or calibrated, in all three applied algorithms. Whereas external acceleration effects were not considered in the first algorithm, external acceleration effects were compensated for in the second and third algorithms, using the Kalman filter’s residual and acceleration model. Low, intermediate, and high external acceleration scenarios were then implemented in our test-bed. Three different rotational frequencies (0.3, 3, and 6 ㎐) for roll and pitch rotations were applied. Performance of each estimation algorithm was analyzed using slopes, y-intercepts, and standard deviations obtained from the linear regression. Our results confirm that attitude estimation errors are linearly proportional to the magnitude of the applied external acceleration. Perhaps most importantly, our results show the second algorithm may be used to provide relatively uniform and accurate estimation performance for low- and high-frequency motions.