ECSE Departmental Research Seminars

Speaker: Therdsak Tangkuampien

E&CSE, Monash University

Abstract:

The ability to realistically capture human motion presents numerous opportunities for real world applications, ranging from human computer interaction interfaces to computer animation and control in movies and computer games. Currently to accurately capture human motions, magnetic or optical markers are systematically attached to an actor and an expensive calibrated system is used to capture the positions of these markers as the actor performs the required motions. The main disadvantages of attaching markers are the restriction imposed on the actor’s motion and the cost of a system specifically designed to track the markers. The aim of this research is to develop a multiple cameras capturing system that will be able to realistically capture human pose without the use of any markers.

The research aims to investigate the possibility of applying non-linear manifold learning techniques like Locally Linear Embedding (LLE) and Kernel Principal Component Analysis to aid in motion capturing. An accurate 3D mesh of the actor, created from point cloud data captured by a laser scanner is used to generate synthetic 3 dimensional representation of the actor in virtual space. A high density set of poses ranging the space of possible human motion is then used to animate the mesh and the resultant images captured by virtual cameras with the same intrinsic and extrinsic parameters as the real-world cameras. Provided that the synthetic image data is well sampled, the set of all possible images of the mesh should lie on a common lower dimensional manifold as the one generated by the set of possible human poses. This intuition is based on the idea that in a constant and controlled environment, the images of an actor captured by cameras with static intrinsic and extrinsic parameters should mainly be dependent on the current pose (joint angles) of the person at that particular frame. Given a new set of real images of the actor, the system can then project the captured image data onto the synthetic common manifold. Once on the manifold, the poses that produce the closest set of synthetic images to the captured images can be determined and used to generate the current output pose of the actor.

About the speaker:

Therdsak (Tk) Tangkuampien is a postgraduate research student in the Department of Electrical & Computer Systems Engineering at Monash University. His research on marker-less human motion capture is conducted under the supervision of A/Prof David Suter and Prof Ray Jarvis.

Tk completed his Bachelor of Science in Engineering (Electrical & Computer Engineering) with University of Cape Town in 2003, and commences his Research Masters in Engineering and Science in June 2004.