Licentiate seminar: Analysis and Optimization of Systems for Detecting Sleepiness in Drivers
David Sandberg, Division of Vehicle Safety, Department of Applied Mechanics, will present his licentiate thesis on Monday September 29th at 10.00.
Title:
"Analysis and Optimization of Systems for Detecting Sleepiness in Drivers"
Discussion leader: Björn Peters, PhD, Research leader at the Swedish National Road and Transportation Research Institute, Sweden.
A light lunch will be served after the seminar. If you wish to participate in the lunch, please inform Sonja Laakso (sonja dot laakso at chalmers dot se)
--------
Abstract
--------
The aim of this thesis has been to generate sleepiness detection systems that can help reduce the number of sleepiness-related vehicle accidents.
The best detection system presented here used a feedforward neural network to combine two sleepiness indicators, and reached a performance score (based on the average of sensitivity and specificity, and with a theoretical upper limit of 1.00) of 0.83 on previously unseen data. This result can be compared with a maximum performance score of around 0.72 obtained for the best indicator (based on a driving behavior signal) taken from the literature.
The detection of driver sleepiness has been carried out using (1) sleepiness indicators, i.e. scalar measures obtained from time series of driving behavior data such as, for example, the lateral position of the vehicle or a mathematical model of sleepiness, and (2) detection systems combining several indicators. Classifiers have been formed by combining sleepiness indicators (or detection systems) with a threshold value.
Thus, driver sleepiness detection was formulated as a binary classification problem (involving the two classes sleepy driving and alert driving) that was attacked using supervised learning methods.
Specifically, the optimization of sleepiness classifiers was carried out using stochastic optimization algorithms, such as genetic algorithms and particle swarm optimization. The optimization procedure was used for setting the parameters of the indicators and detection systems so as to achieve maximum detection accuracy.