Linear Combinations of Upper Body Motion
Yu Huang, Beckman Inst., UIUC
(
) +
c1(
) +
c2(
) +
c3(
) +
c4(
) + ....
c1(
) +
c2(
) +
c3(
) +
c4(
) + ....
c1(
) +
c2(
) +
c3(
) +
c4(
) + ....
Notations: We attached 13 markers to the upper human body (3 on the head, 4 for each arm, 2 on the torso) for tracking by a commercial motion capture system (a six infrared-camera-system, Motion Analysis Corporation). Once the ground truth of 3-D position of each markers was captured, we started learning the motion model by PCA: Sampling from original data, with the overlapped offset of 5 frames, we got 629 10-frame segments, each represented by 390 numbers (13 markers times 3 dimensions times 10 frames). After subtracting a mean vector we did SVD: X=USV', then columns of U are the basis functions (eigenvectors), diagonal elements of S are singular values. The principal eigenvectors will be useful to synthesize a movement or analysis the true movement in the video.