Fast 3-D human motion capturing from stereo data using Gaussian clusters

In our previous work, a new system using a stereo camera has been proposed to estimate human motion from a sequence of 3-D video frames. The system used the articulated human model defined with connected ellipsoids and then co-registered the 3-D human model to 3-D data to recover the 3-D human body posture on each frame. Consequently, the 3-D human motion is reflected by the kinematic angles of the estimated human body postures. However, this approach had a limitation of the prolonged computational time, because, in order to fit the 3-D human model to the 3-D data, a large number of 3-D points were used in the co-registration. In this paper, we have proposed a new co-registration algorithm that uses the Gaussian distribution to cluster of a set of 3-D points into groups for registration. This improvement leads to the reduction of computations and allows our algorithm to be able to process about 10 frames per second and to be more suitable for real time applications.