PULLING out into the fast lane when a juggernaut is bearing down on you is
not one of life鈥檚 best ideas. So it is just as well researchers are developing a
car that can predict when you are about to make a dangerous move and help you
avoid disaster. By observing your behaviour, the car can tell if you are going
to turn, change lanes, brake or pass another car in the next few seconds, and if
need be it will activate appropriate warning or override systems.
By shifting the emphasis of car safety away from the design of the vehicle
itself and looking more towards the driver鈥檚 behaviour, the developers believe
that they can start to build cars that adapt to suit people鈥檚 needs. The
research was a joint project between Andrew Liu at Japanese car maker Nissan and
Alex Pentland at the Massachusetts鈥� Institute of Technology. They say tests
using a driving simulator showed their system to be 95 per cent accurate at
predicting drivers鈥� moves 12 seconds in advance.
Liu and Pentland鈥檚 system relies on the fact that simple driving behaviours
can be broken down into long chains of simpler sub-actions, including
preparatory actions. Even car passengers, after a little thought, should be able
to predict what a driver will do next, say the researchers. To make its
predictions, Nissan鈥檚 smart car uses a computer and sensors on the steering
wheel, accelerator and brake to monitor a person鈥檚 driving patterns. A brief
training session, in which the driver is asked to perform certain manoeuvres,
allows the system to calculate the probability of particular actions occurring
in two-second time segments.
Advertisement
The statistical technique behind the system is similar to that used in speech
recognition software. When driving normally, the driver鈥檚 behaviour is
continually monitored for patterns. When a recognised pattern crops up, the
system predicts the action most likely to follow. And since it continually
updates the probabilities of these actions, it can adapt to the individual
driver, getting more accurate all the time.
Some of Liu鈥檚 earlier work involved tracking eye movements to recognise what
a driver was doing at any one moment. In theory, he says, this predictive system
could easily be adapted to monitor just the driver鈥檚 eyes. This would mean an
even earlier warning system, since the distinctive eye movements a driver makes
when turning, for example, happen well in advance of the actual manoeuvre.
