Towards Multimodal Driver’s Stress Detection - Digital Signal Processing for In-Vehicle Systems and Safety - page 14

Fig. 1.9 Distraction detection algorithm flow based on features extracted from CAN-Bus signals
interval is assigned to a probability. For example, if the ratio is between 0.1 and 1,
the probability of distraction is 0.7, and if the ratio is larger than 20, it is 1.
This assignment approach allows for a probabilistic assessment of the distraction
or can give an idea of the distraction level.
Comparison values larger than 0.1 in magnitude are considered to indicate a
significant distraction. If the comparison value magnitude is below 0.1, the session is
assumed to be close enough to baseline to be considered neutral. As the comparison
ratio increases, the probability of being distracted increases, with the highest value
being 1 as shown in Fig. 1.9 . At the end of this probability mapping, the prob-
abilities are summed along the feature vector (now comprised by comparison ratios)
and normalized by dividing the resultant likelihood value in the feature vector
dimension. The next section explains the feature extraction process and motivation
behind the feature vector elements selected.
1.4.1 CAN-Bus-Based Features
The features are selected based on their relevance to distraction and definition of the
maneuver. Using the color-coded driving timeline plots, it was observed that the
route segment two contains lane keeping and curve negotiation tasks in terms of
driving. For the lane keeping, several driver performance metrics are suggested in the
literature mostly using steering wheel angle (SWA) to calculate a metric indicating
the fluctuations or microcorrections in SWA input. Among these metrics, a widely
accepted method is the sample entropy [ 22 ] and standard deviation. If available,
the lane deviation measurements also give away if the driver is fully attentive and
in control. The reversal rate of steering wheel is also considered to be a reliable metric
to measure driver performance in a lane keeping task. Boer [ 23 ]recentlyupdated
his previous work and suggested some adjustments, taking high-frequency terms