Explicit duration models for isolated hand gesture recognition

In this paper we test the recognition efficiency of explicit duration models (EDM) for isolated gesture recognition. First, through a careful analysis of the characteristics of hand gesture patterns, the shortcomings of homogeneous hidden Markov models (HMM) are pointed out. Next, EDM is proposed as an efficient method to model durations. Finally, to validate these claims, an EDM based framework is developed and tested along with HMMs, hidden conditional random fields and input-output HMMs, on a database consisting of 10 3D hand gestures using 5×2 cross validation. By comparing the testing times of models using parameters that maximize the recognition rates, it is concluded that EDMs are better suited for real-time applications than the other models.