Motion Imitation of a Humanoid Robot via Pose Estimation

Due to the complex operation environments, it is with great significance to improve the flexibility, efficiency and usability of robots while reducing the professional requirements and workload of robot deployment. In this paper, we propose a novel motion imitation approach to a humanoid robot based on 3D human pose estimation. First, we present a sparse-tensor-based 3D human pose estimation approach with point cloud pruning, which solves the dependence of 3D pose estimation on 2D results by directly using the quantized RGB-D data. By effective pruning, the memory stability of 3D data can be guaranteed. Second, a simple yet effective motion retargeting method using coordinate transformation and space vector calculation is applied. Finally, experiments demonstrate the accuracy of our proposed human pose estimation method, thus making motion capture less dependent on expensive and complex motion capture devices and the performance of similarity, continuity and stability of motion imitation is verified.