Digital Signal Processing Reference
In-Depth Information
3.3.2 Driver Identification Model
The ability to identify a driver and his/her driving behaviors is related with how
he/she hits the gas and brake pedals. We model the statistical nature of these pedal-
operation patterns with Gaussian mixture models (GMM). The maximum posteriori
probability approach to the N-class identification problem requires computation
of conditional probability
P
,
N
, given a feature
vector
f
representing the sample data of an unknown class. An alternative is to
employ the maximum likelihood solution, which maximizes the class-conditional
probability,
ð
l
n
j
f
Þ
for each class
l
n
,
n
¼
1,
...
l
¼
arg max
l
n
log
P
ð
f
j
l
n
Þ:
(3.2)
Furthermore, the likelihood scores coming from different classifiers can be
combined at decision level (decision fusion) using weighted summation rule,
X
k
a
k
P
l
¼
arg max
l
n
ð
f
k
j
l
n
Þ;
(3.3)
1 is the weight of the
k-th
classifier and
P
k
a
k
¼
where 0
1.
The computation of class-conditional probabilities needs a prior modeling step,
through which we estimate a probability density function of feature vectors for each
class
a
k
,
N
from available training data. The class-conditional probability
density functions are modeled using the Gaussian mixture densities,
l
n
,
n
¼
1,
...
X
M
k¼
1
o
k
N
P
ð
f
j
l
n
Þ¼
ð
f
;
m
k
;
C
k
Þ;
(3.4)
where
m
k
and
C
k
are respectively mean vector and covariance matrix of the
k-th
mixture, and
M
is the total number of mixtures.
3.3.3 Driver Behavior Prediction
We propose a driver behavior prediction system, which performs temporal clustering
of behavior signals and computes linear estimators for each temporal cluster, based
on the work in [
15
]. The temporal clustering is performed with hidden Markov model
(HMM). Within each temporal segment linear estimators predict current driving
behavior sample from
N
recent samples of all behavior signals. The consistency of
the predicted signal and the actual signal is expected to give us an idea about driving
quality. We employ brake, gas pedal strokes, and velocity for our prediction model.
Flowchart of predicting driver behavior is shown in Fig.
3.5
.
Search WWH ::
Custom Search