Abstract:
Motion capture data is a digital representation of the complex temporal structure of human motion. Motion capture is widely used for data-driven animation in sports,medicine and entertainment, because of its ability to capture complex and realistic
motions. Due to its efficiency and cost, methods for reusing collections of motion capture data are becoming important in the field of computer animation. These motions can then be used for motion blending and morphing, which in turn requires identification and retrieval of the motion from the large collection of motions. Currently, motion data is manually labeled and segmented through a labor-intensive process. This thesis investigates algorithms for the classification of motion capture sequences. This classification task is challenging due to the data being high dimensional, continuous, and time-variant. The main contribution of this thesis is an empirical comparison of a variety of classification algorithms for motion capture sequences. We investigate three different
aspects of these classification algorithms: 1) the use of discrete versus continuous models of the data, 2) generative versus discriminative models and 3) dimensionality reduction through Principal Component Analysis, a linear technique, versus the Gaussian Process Latent Variable Model, a non-linear technique.
