Quasi-Monte and Data-Driven Monte Carlo Methods for Efficient Human Joint Model Fitting

Abstract

Fitting a kinematic model of the human body to an image without
the use of markers is a method of pose estimation that is useful
for tracking and posture evaluation. This model-fitting is challenging
due to the variation in human physique and the large number
of possible poses. One type of modeling is to represent the human
body as a set of rigid body volumes. These volumes can be
registered to a target point cloud acquired from a depth camera
using the Iterative Closest Point (ICP) algorithm. The speed of ICP
registration is inversely proportional to the number of points in the
model and the target point clouds, and using the entire target point
cloud in this registration is too slow for real-time applications. This
work proposes the use of data-driven Monte Carlo methods to select
a subset of points from the target point cloud that maintains or
improves the accuracy of the point cloud registration for joint localization
in real time. For this application, we investigate curvature of
the depth image as the driving variable to guide the sampling, and
compare it with benchmark random sampling techniques.