Producing a neural network for monitoring driver alertness from steering actions

There is a limit to the accuracy with which we can predict a person's state of alertness from their behaviour. Driver behaviour and alertness are, however, clearly related, and this should allow us to build a predictive model. For such a model to be of use it must be very general in its ability. Such generality is available at the expense of accuracy and a trade-off between overall error rate and quantity of usable predictions must consequently be made. This paper discusses a set of methods which were applied to the task of building a neural network based system for predicting driver alertness from steering behaviour. We show how an acceptable level of generality was achieved and how the trade-off between error rate and quantity of usable predictions was managed.