Articulated Human Body Pose Inference from Voxel Data Using a Kinematically Constrained Gaussian Mixture Model

We present a novel method for learning and tracking the pose of an articulated body by observing only its volumetric reconstruction. We propose a probabilistic technique that utilizes a multi-component Gaussian mixture model to describe the spatial distribution of voxels in a voxel image. Each component describes a segment or rigid body, and the collection of components are kinematically constrained according to a pre-specified skeletal model. This model we refer to as a kinematically constrained Gaussian mixture model (kc-gmm). Pairs of components connected at a common joint are encouraged to assume a particular spatial configuration, forming a 1, 2 or 3 degree-of-freedom (DOF) joint. The pose learning algorithm, based on the EM algorithm, is evaluated using synthesized hand data, and the HumanEvaII dataset for facilitating algorithm comparison among different algorithms. Both datasets contain groundtruth information for accuracy measurements. A 16 component, 27 DOF mixture hand model and an 11 segment, 19 DOF mixture human body model were used. The results show that utilizing volume data, aided only by the degreesof-freedom constraints, show accuracies of joint location estimates within 0.5cm mean-absolute-error from groundtruth with the hand data set and 17cm from subjects S2 and S4 from the HumanEvaII datasets.