Abstract
The variety of driving styles in which driving manoeuvres are performed and the uncertainty present in the available sensor data are major obstacles to automatic manoeuvre recognition in the automobile. A Bayesian model of the driving task is used for the probabilistic inference of the manoeuvre being performed from the uncertain evidence available. Top-down inference about the driver's likely intentions is combined with bottom-up inference of the manoeuvre from evidence about control actions and vehicle behaviour. A prototype manoeuvre-recognition system identified the correct manoeuvre during 86.5% of the duration of two trial drives on public roads. For the covering abstract see ITRD E134653