We present two data-driven importance distributions for particle filterbased
articulated tracking; one based on background subtraction, another on depth
information. In order to keep the algorithms efficient, we represent human poses
in terms of spatial joint positions. To ensure constant bone lengths, the joint
positions are confined to a non-linear representation manifold embedded in a
high-dimensional Euclidean space. We define the importance distributions in the
embedding space and project them onto the representation manifold. The resulting
importance distributions are used in a particle filter, where they improve both
accuracy and efficiency of the tracker. In fact, they triple the effective number of
samples compared to the most commonly used importance distribution at little
extra computational cost.