Adaptive Directional Walks For Pose Estimation From Single Body Depths

In this paper, we introduce a novel body pose estimation method
based on single depth images with our proposed random forest classifier,
whereby it is possible to estimate the positions of joints
directly with significant accuracy.
We train randomized classification trees based on a joint entropy
objective function combined with the geodesic distances and directional
vectors simultaneously, to estimate the probability distribution
for the label of the directional vector towards a particular body joint
by considering the geodesic structural information. At a test step,
an arbitrary point moves as much as the magnitude
of the probability predicted adaptively by following the predefined kinematic
graph, which is referred to as adaptive directional walks. Numerical and visual
experiments with real datasets confirm the usefulness
of the proposed method.