Driver Status Identification from Driving Behavior Signals - Digital Signal Processing for In-Vehicle Systems and Safety - page 46

Fig. 3.14 State sequence of a test data

The MMSE estimation for each temporal segment is performed under two
different scenarios. In the first scenario, we use direct samples of behavior signals
that are actual recent readings from sensors, at each estimation step. In the second
scenario, within each temporal cluster we start using recent actual samples but
continue with the estimated samples to estimate upcoming samples. Hence, in the
second scenario, estimation looks forward based on the current actual signal
readings. The second scenario creates a more realistic system to foresee expected
driving behavior characteristics.
The driving behavior signals are predicted using a window of past behavior
samples. All behavior signals are decimated by four for the driving behavior
prediction experiment. The cepstral features for the HMM clustering are extracted
over 800 ms windows (25 samples) for every 96 ms frames (three samples). First,
we build a temporal correlation between all three signals using HMM structure
shown in Fig. 3.14 . This structure is specified by the following parameters:

• Set of discrete states S ¼ {S i ,i ¼ 1, 2, ... , M }

• State transition probability a ij , i ¼ 1, 2, ... , M; j ¼ 1,2, ... , M

where M denotes the number o f states. Transition probabilities from one state to
another are set equal initially, and system starts with probability 1 at the first state.
The HMM model is then trained using EM algorithm with the training data. In the
testing phase, the Viterbi decoding algorithm determines the state sequence of the
test data. A sample state sequence using eight-state HMM clustering is given in
Fig. 3.15 .
Figures 3.16 and 3.17 , respectively, present samples of driving behavior signal
prediction based on the first and second scenarios for a randomly selected driver.