7.2.2 Long-term work

We now consider major aspects of research that could carry on with the work of this thesis.

• Dealing with the segmentation issue. A major problem for computer-based learning is segmenting continuous-time data into its structural components. Sign language is no different. Breaking up continuous Auslan into its component signs is very difficult. It is expected that closely related to this would be the creation of a real-time continuous Auslan recogniser, which would be the next step.

• Investigating other approaches that make use of the features. There is no reason why some of the features used here could not be used as the input to a hidden Markov model. In particular, using the HMM would allow us to simultaneously address the problem of continuous Auslan. Similarly a probabilistic grammar approach might be just as effective.

• Combining the data with facial gesture information. We could also include information about what the face is doing as well, since this would provide a useful clue to the nature of the sign. Already, some work has been done on facial gesture recognition.

• Work on finding suitable transformations for new users of a system such as GRASP to a common space where the concept boundaries are signer independent; or developing more signer-independent features. This can be viewed as a highly complex issue in calibration.

• Use contextual information to aid in the recognition of signs by using techniques such as trigrams, bigrams, sign grammars and so on.

Research in this area is by no means limited to the above and there is much, much more scope for research. This area is interesting not just because of the challenges it poses to learning and to data analysis techniques, but also because of its potential application for creating a new channel of communication between humans and computers that can be far more intuitive than a keyboard.