Detecting Sleepy Drivers by Pattern Recognition based Analysis of Steering Wheel Behaviour

摘要

Using steering wheel behaviour based approaches for sleepiness monitoring might have the advantage of being cheap, non-intrusive, and robust even under extreme demanding environmental conditions (e.g. high background noise, temperature, or humidity). Twelve healthy young adults completed 7 overnight driving sessions (1 - 8 a.m.) in our real car driving simulation lab. The combinations of observed and self-rated sleepiness (Karolinska Sleepiness Scale; KSS) were considered as ground truth of sleepiness. Steering angle, lane deviation and pedal movement were recorded. In order to investigate sleepiness-induced changes in steering wheel behaviour, spectral and state space domain features are computed. Using advanced signal processing procedures for feature extraction, we computed 3 feature sets in the time, frequency and state space domain (a total number of 1251 features) to capture fatigue impaired steering patterns. Within the time domain we extracted class distribution measures, peak amplitudes and distances, and zero crossing distances (e.g. maximum of peak amplitude, mean distance between consecutive zero crossings). Each feature set was separately fed into 5 pattern recognition methods (e.g. SVM, KNN, MLP). We yielded a recognition rate of 86.1% in classifying slight from strong fatigue.