A complete scheme for motion prediction based on motion capture data is presented. The scheme rests on three main components: a special posture representation, a diverse motion capture database and prediction method.
Most prior motion prediction schemes have been based on posture representations based on well-known local or global angles. Difficulties have arisen when trying to satisfy constraints, such as placing a hand on a target or scaling the posture for a subject of different stature. Inverse kinematic methods based on such angles require optimization that become increasingly complex and computationally intensive for longer linkages. A different representation called stretch pivot coordinates is presented that avoids these difficulties. The representation allows for easy rescaling for stature and other linkage length variations and satisfaction of endpoint constraints, all without optimization allowing for rapid real time use.
The validity of this scheme also rests on the availability of motion capture data. There are two situations - in one case the user has access to a larger database of motions relevant to the particular problem while in the second case, the user collects a small amount of motion capture data concerning the task of interest. At the Human Motion Simulation Laboratory (HuMoSim) at the University of Michigan, we have collected several large databases on various types of automobile and materials handling motions. A prediction models based on these databases is presented.
Two contrasting prediction methods are demonstrated. One is parametric and uses functional regression analysis to predict the stretch pivot coordinates used in the postural representation as they vary over the time of the motion. This regression-type model allows the use of subject-based variables such as stature and age and task based variables such as target location and object weight to influence the predicted motion. This also allows the scientific study of the effect of such factors using statistical significance testing. The second type of prediction method is based on the idea of nearest neighbor nonparametric regression. A small number (perhaps even just one) of motions is selected that have characteristics similar to the motion we wish to predict. These motions are then averaged in a special way using the stretch pivot coordinates to produce the predicted motion with the required features such as stature, target location etc. This method lends itself to the prediction of motion based on small special purpose motion capture databases collected by the user for some specific problem.