We seek to extract the three-dimensional motion of articulated objects by passive sensing. We first developed a multi--view system that segments objects from the background before their subsequent analysis by a motion model. To reduce the effects of noise, we also performed 3D reconstruction prior to motion analysis to improve the extraction of the (static) background. The 3D extraction and motion analysis are both performed by evidence gathering, thus accruing optimality of performance.
These new analyses are supported by a new representation which combines resolution with depth thus improving on a conventional voxel representation. We show how we can extract and represent moving bodies by applying these new techniques to video rate imagery of a walking human subject, and of a ball moving under the influence of gravity, and describe their associated advantages.
