Online optimal path decoder of hidden Markov model and its application to connected gesture recognition

We model a recognition problem for connected hand gestures to find an optimal path through a hidden Markov model (HMM) directed acyclic graph. To determine this optimal path, an online graph search method is proposed that decodes the observed gesture pattern and evaluates the optimal graph node at each time frame of the continuously deepening HMM graph. The temporal characteristic of gesture recognition is subsequently handled by introducing a rejection threshold time that acts as a depth-wise sliding window for pruning unnecessary graph nodes. The functional depth of the graph is defined by this depth rejection threshold. Experimental comparison of our algorithm with other HMM-based search algorithms demonstrates the effectiveness and robustness of our method. (c) 2008 Society of Photo-Optical Instrumentation Engineers.