Driver Status Identification from Driving Behavior Signals - Digital Signal Processing for In-Vehicle Systems and Safety - page 43

Fig. 3.11 Histograms of gas pedal pressure signals under reference driving (top left), dialog on
cell phone (top right), signboard reading (bottom left), and dialog with passenger (bottom right)
cepstral features are calculated for every 1-s frame. Figure 3.11 shows the
histograms of the gas pedal pressure signal under different driving task conditions.
Statistical differences between reference driving and driving under a task are
observable. However, driving-task-specific histograms, especially dialog on cell
phone and with passenger, are close to each other.
We first consider a two-class classification system to identify reference driving
and driving with a task. The two-class identification system is expected to show
whether distractive conditions are influential on driving performance. Among
U-TASK dataset reference, driving lasts 190 min (47.8% of all data) and driving
under a task lasts 207.5 min (52.2% of all data) totally. To evaluate the task
identification, we use 16-mixture Gaussian classifiers and fivefold cross validation
for classification.
The average identification rates of the classifiers using gas and brake pedal
pressure signals, and their decision fusion with different decision-window sizes
are given in Fig. 3.12 . The best scenario is achieved by using 360 s of decision
windows.
Table 3.1 shows the identification rates of each class for this scenario. In this
table, the last column presents the prior reference distribution of events in the
database. We identify a reference driving session with 93.2% of success and
identify a driving session under a specific task with 72.5% of success by using
the fusion of gas and brake pedal signals. The average task vs. no-task identification
result is obtained as 83.3% with the fusion of 16-mixture GMM classifiers of gas
and brake signals. Note that, these identification rates are significantly higher than
random classifier performances with possible uniform distributions.
We also consider identification of individual tasks from driving behavior signals.
Among all driving sessions under a specific task, dialog on cell phone lasts 97.5 min