Research Seminar - March 12, 2002

Automating tracking and reconstruction of human motion

Abstract:

Human motion is mainly defined by the movements of underlying skeleton. This skeleton can be modelled as an articulated structure of rigid links connected at rotational joints forming a kinematic chain. Therefore, human movement can be summarised by the motion of these rotational joints, and the fixed lengths of the rigid links. This work involves automating the tracking of the projection of these joints, 3D reconstruction of the motion, as well as outlier rejection of tracked data. When tracked data is rejected, the missing data must be inserted into the 3D structure. This is accomplished by using the metric structure of the kinematic chain. These inserted points can then be backprojected into the images, in order to iteratively improve the tracking. The results presented are of self calibration and 3D reconstruction of a tennis stroke from two views. Contrary to model based systems, neither precalibrated cameras were required, nor manual intervention for initialisation and error recovery.