Drivers falling asleep behind the steering wheel are the cause of many
traffic accidents, and the statistics show that the number of sleepiness
related accidents are escalating. Commercial drivers represent a large part
of the sleepiness accident statistics, probably depending on much time spent
on the road, long driving hours and the monotonous character of the roads
traveled. Systems for sleepiness detection exist but the evidence to judge
their applications and performance is inadequate. Sleepiness detection from
cameras monitoring the driver and other driver related measures can be hard
and expensive to implement. A system only using variables that could be
measured from the vehicle itself, preferably using already existing sensors,
would be desirable.
The assignment of this master thesis project, commissioned by Scania CV AB
in Södertälje, was to investigate the possibility to develop an algorithm
that detects a sleepy driving behavior, using in-vehicle variables only.
This project is a continuation of a previous master thesis project that
investigated a patent claiming to be able to detect inattentive driving. The
authors came to the conclusion that two of the variables in the patent
showed promising results that should be further investigated. These were to
be tried out in this project, along with other variables proved to predict
driver sleepiness, by performing extensive tests.
Quantitative testing, where 22 subjects drove a simulator while sleep
deprived, enabled the collection of ten raw variables, measured from either
the steering wheel or lane position. Examples of raw variables are steering
wheel torque, yaw angle rate and lateral acceleration. These were combined
in different ways to form 17 transformed variables that according to
literature had shown to be correlated with a sleepy driving behavior, like
the number of lane exceedances or the variance of lateral position. To be
able to judge the performance of the different transformed variables, a
reliable measure of the driver's actual sleepiness was needed. A subjective
measure called Karolinska Sleepiness Scale (KSS) was chosen, where the
drivers estimate their own sleepiness on a 1-9 scale. The best version of
each transformed variable was optimized compared to the KSS and forward
selection with regression analysis was used to extinguish which variables
should be combined to make the best formula to detect sleepiness. Since some
transformed variables were not defined for all time intervals, different
formulas had to be created depending on which variables that was available.
This created a selection model where six different formulas were used.
The algorithm performance was judged and it proved to give good results. The
formulas combined in the algorithm make correct classifications, sleepy or
alert driver, in more than 87 % of the cases when sleepiness threshold was
set to eight, with a low false alarm rate of less than one percent. This is
a promising result considering that only in-vehicle variables were used. A
better performance would probably come from combining the detection from in-
vehicle variables with another sleepiness measure.
The project is done in collaboration Jens Berglund from Linköping
University. The work was divided during the literature study ...
