Next: 6 Conclusions Up: Using Hidden Markov Models Previous: 4 Learning gesture models

5 Gesture classification

Since the HMM with continous-observations allows for several levels of description, it is suited to describe a generic type of gesture, not trivially divisible in a small number of signs. Gesture recognition becomes a classification problem. The comparison between the new image sequence and each memorized model of recognizable gestures consists in updating the parameters of the model with the new sequence. If the model fits well the new observation sequence the learning procedure must show an essential stability of the likelihood function, see Figure 11; otherwise the model is rejected.

Figure 9: Final HMM generated by the learning process with N =4. Each state corresponds to the canonical posture illustrated in its neighbouring picture.

Figure 10: The likelihood function shows that the learning process converges in few steps.

Next: 6 Conclusions Up: Using Hidden Markov Models Previous: 4 Learning gesture models

Adrian F Clark
Mon Jul 28 12:54:58 BST 1997