Motion capture is a generic term denoting recording movement and translating that movement onto a digital model. In particular application to the human body, the purpose of motion capture is to measure the motions of rigid segments of the human body during various activities. Such systems have numerous applications, including human motion study in biomechanics, spanning from clinical gait analysis to sport performances and injury prevention, military applications, and motion picture industry. In filmmaking, motion capture (also known as mocap) is used to record actions of human actors, and using that information to animate digital character models in 2D or 3D computer animation.
Most existing motion capture systems track special markers placed on rigid parts and joints of the body. The disadvantage of such approaches is first of all the fact of objects attached to the human body or clothes, which in some cases make the subject feel uncomfortable and make the motion unnatural. Secondly, since typically a small number of markers are used, only a sparse set of points is available for tracking. Subtler motions, such as non-rigid deformations of the soft tissues are lost in this way. Finally, the use of markers limits the application to controlled scenarios, in which the captured object is prepared and well-known in advance.
The scope of the proposed project is to develop a markerless motion capture approach for analysis of generic dynamic objects in natural conditions. Besides removing the standard drawbacks of marker-based motion-capture systems, markerless motion capture would constitute a qualitative improvement, and it would no more be limited to a calibrated lab setting, but could rather be applied in a wide range of natural scenarios. Of particular interest are applications relevant to autonomous systems, such as robot navigation and interaction with surrounding objects (e.g. humans). The proposed solution would naturally allow motion capture of multiple objects and would not be limited to objects known in advance. Motion capture in autonomous systems could become important component, adding novel functionality and extending existing one.
The proposed system will be based on a 3D sensor that acquires the body geometry in motion, producing a 3D video of a moving subject. At the second stage, a correspondence between the 3D video frames is established. Since the body is non-rigid, we use intrinsic correspondence which is deformation- invariant. Given the correspondence between every two consecutive 3D frames, the motion is estimated using a 3D version of over-parameterized optical flow. This process automatically determines rigid parts of the body. Finally, the recovered motion and articulated parts will be fit into a kinematic model.
