Fast Skeleton Estimation from Motion Capture Data using Generalized Delogne-Kåsa method

Abstract

This paper presents a fast closed-form solution estimating the rotation points of joints relative to the motion capture data. The proposed solution estimates the physical location of joints inside the body of the person wearing the trackers. The Generalized Delogne-Kåsa method used for our implementation fits spheres, cylinders, circles and planes to the motion capture data without the need for initial guessing. This non-iterative, closed-form solution is fast as it calculates the rotation point with O(N) averaging along with one inversion of a 3x3 positive semi-definite matrix for each joint. The error in the joint location is on average 3σ/√N which is low. In addition, sample points for every joint can be from different time sequence allowing flexibility in recovering the joint locations. Once the joint location relative to the tracker position is determined, it could be used for the remainder of the data set. Publicly available CMU motion capture data was used for this study. Two animation sequences, showing our method, are included with this paper. These results can be compared to that available at the CMU site for the same animation. Since the pose is found relative to the given data, our pose estimation provide better fit to the given data, revealing subtle, individual nuances of the person used for the motion capture. Because of the closed form solution, our technique is ideally suited for the use of motion captured data to create skeletal motion in 3D games or applications where real time performance is essential.