Multilayer Modeling of Driver Behavior Based on Hierarchical Mode Segmentation - Digital Signal Processing for In-Vehicle Systems and Safety - page 57

Fig. 4.2 Transformation from data space to feature vector space

where y LDs j
l

( c 1; l ¼ 1, 2) is the output samples in the LDs j , and F j is given by

T

F j ¼ð

r 1 r 2 r c Þ ð r 2 LDs j Þ : (4.8)

As the result,

LD j; 1 T LD j; 2 T

m j T

T

x j ¼ððy Þ ; ðy Þ ; Þ (4.9)

:

2. For each feature vector x j , the following covariance matrix R j is calculated:

0 1

V j; 1 00
0 V j; 2 0
00 Q j

@ A

R j ¼ (4.10)

where

SSR j;l

F j Þ 1

Þ ð F j T

V j;l ¼ (4.11)

c ð 9 þ 1

y LDs j
l

y LDs j
l

T

F j ð F j T

F j Þ 1 F j T

SSR j;l ¼ ð I Þ (4.12)

X
r2LDs j ð

T

Q j ¼ r m j Þð r m j Þ (4.13)

:

x j represents the combination of the local dynamics and data.
By this definition, the data is classified based not only on the value of data but
also on the similarity of the underlying dynamics. Furthermore, the covariance
matrix R j represents the confidence level of the corresponding feature vector

The feature vector

x j .
R j is used as the weighting matrix in the calculation of the dissimilarity between
feature vectors in the clustering procedure.