Motion Models for Monocular 3--D Human Body Tracking
Click on the images to jump
to some results.
We investigated the use of motion models based on either Principal Component
Analysis or Gaussian Processes to formulate the monocular tracking problem as
one of minimizing differentiable objective functions. As shown in the
results below, we can handle cyclic motions, such as walking
and running, as well as acyclical ones, such as a golf swing.
Furthermore, by detecting key postures, we can achieve robust and fully
automated 3D motion recovery of people seen from arbitrary viewpoints by a
single and potentially moving camera.
By tracking the feet, knees, hands, and head
of the golfer in the video sequence and fitting a parameterized model of a
golf swing to the resulting trajectories, we can recover the 3D motion of the
skeleton.
On the left we see the 3D skeleton
reprojected onto the original video sequence and on the right a 3D avatar that
performs the same motion as the real golfer. The hands were tracked using
a method we developed
in earlier work.
The algorithm automatically detects the pedestrian at the moment of the walking cycle when his legs are furthest apart and fits a walking motion model in between detections. As a result the subject can be tracked from an arbitrary viewpoint, his full 3D motion recovered, and his instantaneous speed estimated.
