Driver Status Identification from Driving Behavior Signals - Digital Signal Processing for In-Vehicle Systems and Safety - page 35

Fig. 3.4 Driving behavior signal histograms of two drivers: driver one on the left, driver two on
the right columns

separation method, and it has been used for driving behavior signals. Cepstral analysis
captures significant information from driving behavior signals. In driver modeling,
hitting a gas or brake pedal is filtered with a driver model represented as the spectral
envelope. Spectral envelopes of pedal-operation signals represent the differences in
pedal-operation patterns. These spectral envelopes are similar in the same driver and
vary across different drivers.
In this study, we extract cepstral features for the gas and brake pedal pressure
and velocity signals, which are sampled at 32 Hz. The cepstral features are
extracted over 800-ms windows for every 96-ms frames. The cepstral feature is
defined as the first K coefficients of the discrete cosine transform of band-pass
filtered log-magnitude spectra,

f k ¼ DCT f BPF f log j F f x w ð n þ kT Þgjgg (3.1)

where k is the frame index, x w ð is the windowed signal of duration T .

In order to eliminate high-frequency noise, we apply band-pass filtering with
1-13 Hz cutoffs for brake signal and with 1-6.5 Hz cutoffs for gas and velocity
signals. The dimension of the feature vector is set as K

n þ kT Þ

¼ 10.