Towards Multimodal Driver’s Stress Detection - Digital Signal Processing for In-Vehicle Systems and Safety - page 15

Table 1.4 Feature vector and definitions
Notation Definition

WDE_SWA Wavelet decomposition detail signal energy for SWA

WDE_speed Wavelet decomposition detail signal energy for speed

SampEnt_SWA Sample entropy of SWA

SampEnt_speed Sample entropy of SWA

STD_SWA Standard deviation of SWA

STD_speed Standard deviation of SWA

STD_SWAR Standard deviation of SWA rate

into account. It was also pointed out in a thorough analysis [ 24 ] that the speed interval
for which the SWA-dependent metric is being calculated is important since the lower
speeds require more SWA inputs to achieve the same amount of lateral movement of
the car compared to a higher speed. For the curve negotiation, a constant input of an
angle required using the visual input of the road curvature.
The novice or distracted driver may have fluctuating inputs in the SWA, and the
general trend is that the speed should be reduced while taking the curves to balance
the centrifugal force. Although different in nature, lane keeping and curve negotia-
tion can be seen as regulatory control tasks from the driver's point of view.
Therefore, we selected a seven-dimensional feature vector using available informa-
tion and observations about driver performance/behavior including: energies of
high-frequency components wavelet decomposition (WD), sample entropy, standard
deviation, and standard deviation of rate of change (R-STD). All features are
extracted for SWA, and speed channels except R-STD are only applied to SWA.
The time window length is taken as equal to the maneuver length, and the effect of the
signal length is eliminated in the calculation of features. The entries of the feature
vector are listed with their definitions in Table 1.4 .
For the wavelet decomposition, Daubechies [ 25 ] wavelet kernel with fourth
order is used, and detail signal is taken at the sixth level. Daubechies wavelet is
chosen since it can approximate to signals with spikes and discontinuous attributes
well. The level and order is adjusted to be able to extract the high-frequency content
in the signal which is in the limitation of human control; the higher details are
ignored since they might be caused by other disturbances in the measurement rather
than driver. Scaling functions (a), wavelet function coefficients (b), scaling function
(c), and wavelet function (d) for DB4 are given in equation group (1.6):

p 3 p 3 p 3 p 3

1 þ 3 þ 3 1

2 p ; 2 p ; 2 p ;

p ;

h 0 ¼ h 1 ¼ h 2 ¼ h 3 ¼ (1.6a)

4 4 4 4

g 0 ¼ h 3 ; g 1 ¼ h 2 ; g 2 ¼ h 1 ; g 3 ¼ h 0 ; (1.6b)

a i ¼ h 0 s 2 i þ h 1 s 2 iþ 1 þ h 2 s 2 iþ 2 þ h 3 s 2 iþ 3 ; (1.6c)

c i ¼ g 0 s 2 i þ g 1 s 2 iþ 1 þ g 2 s 2 iþ 2 þ g 3 s 2 iþ 3 : (1.6d)