Human Motion Tracking from Video Mitsubishi Electric Research Lab Overview We developed a system for fitting a 3D human model to a single video sequence (such as the video on the left) without the need for markers. Applications of human motion tracking include gesture recognition and computer animation (right). Learning 3D Human Motions The human model is learned from a set of motion capture training data, where markers are placed a the person. The markers data is rendered (far left) and a simple cylinder model is fit (near left). Novel motions are expressed using linear combinations of the training motions. 3D Motion Reconstruction from Video The movie above, left shows a Barishnykov dance sequence. The human model was fit to the video, with a limited amount of human interaction. Three views of the 3D model are also shown. Even though there is only one view, the system induces the depth from the prior motion model. Patent "System for reconstructing the 3-dimensional motions of a human figure from a monocularly-viewed image sequence."         Bill Freeman and Michael Leventon.         Patent # US-6,115,052