Signal Processing - November 2016 - 17

Driver 3
Gas Pedal Operation
Patterns with
Rectangular Pulses

Gas Pedal Position (%)

Driver 2
Adjusts Gas Pedal
Position Frequently

Gas Pedal Position (%)

Driver 1
Exhibits Smooth Gas
Pedal Operation
Patterns

Gas Pedal Position (%)

76.8% for 276 drivers when using cepstral features extracted
pedal-operation behavior can be observed. For example,
from pedal-operation signals, compared to 47.5% when raw
Driver 1 operated the pedals more smoothly and maintained
pedal-operation signals were used.
the same pedal positions longer, Driver 2 adjusted pedal positions frequently, while the pedal position patterns of Driver
3 exhibited rectangular pulses. These drivers maintained
Modeling car-following behavior using GMMs
consistent pedal-operation patterns, regardless of the vehicle
We extended our GMM-based behavior modeling framework
being driven.
to predict car-following behavior [13]-[16]. This enabled us
To model such driver's individual characteristics of
to generate car-following behavior patterns for individual
pedal-operation behavior, we applied cepstral analysis and
drivers, allowing for the personalization of headway control,
GMMs to the recorded pedal-operation
which represents a driver's preferred carsignals [12]. We assumed that the impulse
following distance. Gas- and brake-pedal
To model such driver's
signal of a driver's intention to depress
operation patterns of the target driver were
individual characteristics
the gas or brake pedal is filtered by driver
modeled using GMMs as a joint density
of pedal-operation
characteristics, which can be represented
with other driving signals. Gas- and brakebehavior, we applied
as a spectral envelope. The output of this
pedal operation signals while car followcepstral analysis and
system is observed as a pedal-operation
ing were generated from a joint probability
GMMs to the recorded
signal. GMMs were used to represent
distribution of the features modeled with
distributions of the spectral characterisGMMs so that conditional probability was
pedal-operation signals.
tics of pedal operation, extracted through
maximized for the other observed driving
cepstral analysis of the raw pedal-operation signals. The
signals. We confirmed that the GMM-based driver models
resulting driver-behavior model was evaluated through
could replicate driving signals representing the individual cardriver identification experiments using on-road driving data
following patterns of different drivers [13]. We then proposed
collected with an instrumented vehicle, as described prea driver-model adaptation scheme that allowed us to enhance
viously [10]. Experimental results showed that the GMMthe model's capability to represent the particular driving charbased behavior model achieved a driver identification rate of
acteristics of individual drivers [14], [15].

Vehicle A
(Sedan, GV)

50
40

Vehicle B
(Hatchback,
EV)

Vehicle C
40 (Mini-Van, GV)
30

0
10
20
Time (Seconds)
Vehicle A

50
40

0
10
20
Time (Seconds)

Vehicle B

0
10
20
Time (Seconds)
Vehicle A

50
40

0
10
20
Time (Seconds)

Vehicle B

0
10
20
Time (Seconds)

0
10
20
Time (Seconds)
Vehicle C

0
10
20
Time (Seconds)

FIGURE 2. Individual differences in gas-pedal operation patterns.
IEEE SIgnal ProcESSIng MagazInE

November 2016

Table of Contents for the Digital Edition of Signal Processing - November 2016