Three-dimensional motion capture for human sub jects is underde- termined
when the input is limited to a single camera, due to the inherent 3D ambiguit y of 2D video. We present a system that re- constructs the 3D motion of
human sub jects from single-camera video, relying on prior knowledge ab out
human motion, learned from training data, to resolve those ambiguities. After
initializa- tion in 2D, the tracking and 3D reconstruction is automatic; we show
results for several video sequences. The results show the p ower of treating 3D
b o dy tracking as an inference problem.