Abstract:
A visual perception algorithm of human motion for humanoid robot is proposed to improve the precision of the motion data captured from a Kinect. Firstly, the positions of the joints are transformed into the angles according to the inverse kinematics equations. Secondly, the long-time motion is segmented into episodes automatically based on the change of the angular velocity and acceleration, and then RVM (relevant vector machine) is utilized for estimating the angle trajectories with high accuracy. Finally, the spatial consistence, temporal consistence and smoothness of the angle trajectories are given to evaluate the algorithm, and a motion data series processed by the algorithm is implemented on a NAO robot platform. The experimental results indicate that the proposed algorithm effectively improves the spatial and temporal consistence of the motion perception and the smoothness of the trajectory, which provides a foundation for high-precision motion imitation.