Parametric HMMs for Movement Recognition and Synthesis

A common problem in human movement recognition is the recognition of movements of a particular type (semantic). E.g., grasping movements have a particular semantic (grasping) but the actual movements usually have very different appearances due to, e.g., different grasping directions. In this paper, we develop an exemplar-based parametric hidden Markov model (PHMM) that allows to represent movements of a particular type. Since we use model interpolation to reduce the necessary amount of training data, we had to develop a method to setup local models in a synchronized way.
In our experiments we combine our PHMM approach with a 3D body tracker. Experiments are performed with pointing and grasping movements parameterized by their target positions at a table-top. A systematical evaluation of synthesis and recognition shows the use of our approach. In case of recognition, our approach is able to recover the movement type, and, e.g., the object position a human is pointing at. Our experiments show the flexibility of the PHMMs in terms of the amount of training data and its robustness in terms of noisy observation data. In addition, we compare our PHMM to an other kind of PHMM, which has been introduced by Wilson and Bobick.