Comparison of approaches to continuous hand gesture recognition for a visual dialog system

Continuous hand gesture recognition requires the detection of gestures in a video stream and their classification. In this paper two continuous recognition solutions using hidden Markov models (HMMs) are compared. The first approach uses a motion detection algorithm to isolate gesture candidates followed by a HMM recognition step. The second approach is a single-stage, HMM-based spotting method improved by a new implicit duration modeling. Both strategies have been tested on continuous video data containing 41 different types of gestures embedded in random motion. The data has been derived from usability experiments with an application providing a realistic visual dialog scenario. The results show that the improved spotting method in contrast to the motion detection approach can successfully suppress random motion providing excellent recognition results.