Tech Report CS-03-08

Video-Based Tracking of 3D Human Motion Using Multiple Cameras

Ben Sigelman

May 2003


We present a system designed to track a human in three dimensions given video sequences from multiple cameras. The tracking problem is formulated as a Bayesian inference task, and we approximate the posterior distribution with a particle set. When many particles share very little weight, we find that tracking results suffer. We try to prevent such a circumstance by introducing a dynamic mapping from measurements to likelihoods. Additionally, we found that multiple cameras stabilized tracking in three dimensions by disambiguating limb locations in the depth dimension for any one specific camera.

(complete text in pdf or gzipped postscript)