A Non-Parametric HMM Learning Method for Shape Dynamics with Application to Human Motion Recognition

The shape dynamics, i.e., the spatial-temporal shape deformation of an object during its movement, provides much important information about the identity of the object, and even motions performed by the object. In this paper, we proposed a system recognizing object motions based on their shape dynamics. In the proposed system, we use Kenall's definition of shape to represent the object contour extracted from each frame, and construct a tangent space with the full Procrustes mean shape as the pole to approximate a linear space for the dataset, in which the Euclidean distance metric can be used to approximate the full Procrustes distance between shapes. The spatial-temporal shape deformation in motions is captured by hidden Markov models. Since in the traditional HMM framework the hidden states are typically coupled with the training data, which will bring many un-desired problems to the learning procedure, we introduce a non-parametric HMM approach that uses continuous output HMMs with arbitrary states (decoupled from training data) to learn the shape dynamics directly from large amounts of training data where a non-parametric kernel density estimation algorithm is applied to learn the observation probability distribution in order to compensate for the uncertainty introduced by those arbitrary hidden states. This optimizes the HMM training procedure. We then use the proposed system for view-dependent human motion recognition