Human Body Modeling and Tracking using 3D Voxel Data from Infrared Images

We present a system to obtain real-time tracking data of a subject from multiple synchronized Infrared video streams and use this data to drive an animated avatar. Six calibrated infrared cameras are used. Acquired video frames are segmented through background subtraction to obtain the foreground silhouette. Voxel-silhouette intersection tests are performed. The voxels whose projections lie outside the silhouette are discarded leaving behind a carved body model of the subject present in the tracking region. Ellipsoids whose orientation and position are described based on a twist-based framework are fitted onto the voxel model. Based on the inverse kinematic equations for joint angles, the centroid positions of each ellipsoid is calculated using data from 23 measurement points on the 3D voxel model.