Modeling Natural Human Hand Motion for Grasp Animation

Sammanfattning: This report was carried out at Gleechi, a Swedish start-up company working with implementing hand use in Virtual Reality. The thesis presents hand models used to generate natural looking grasping motions. One model were made for each of the thirty-three different grasp types in Feix’s The GRASP Taxonomy. Each model is based on functional principal components analysis which was performed on data containing recorded joint angles of grasping motions from real subjects. Prior to functional principal components analysis, dynamic time warping was performed on the recorded joint angles in order to put them on the same length and make them suitable for statistical analysis. The last step of the analysis was to project the data onto the functional principal components and train Gaussian mixture models on the weights obtained. New grasping motions could be generated by sampling weights from the Gaussian mixture models and attaching them to the functional principal components. The generated grasps were in general satisfying, but all of the thirty-three grasps were not distinguishable from each other. This was most likely caused by the fact that each degree of freedom was modelled in isolation from each other, so that no correlation between them was included in the model.