IEEE Robotics & Automation Magazine - December 2022 - 119

Texture Discrimination
To qualitatively realize texture discrimination, an SVM algorithm
was used to create a classifier for different surface textures.
To consider a range of surface roughness and material
hardness, four surface textures-writing paper, tissue paper,
sandpaper, and flannelette-were used in the present work.
Specifically, by scanning an object while maintaining a fixed
distance between the whisker's base and object's outermost
contour, the robot can obtain a set of raw data containing texture
information. To establish a training data set for the SVM
algorithm, we collected 40 sets of data for each texture surface
under the same conditions. Afterward, the data were transformed
into the frequency domain, where more texture features
can be obtained.
The relevant features for use in the SVM should be specifically
selected while taking generalization performance, computational
efficiency, and feature interpretability into account
[19]. Therefore, we extracted four different features to represent
the characteristics of the original signal: the signal energy
(ENG), spectral entropy (SEN), spectral centroid (SCE), and
average interval of peaks (AIP). The main reasons for using
these features are as follows. First, ENG is a good indicator of
the range of vertical motion experienced by the whisker,
which is typically large for uneven surfaces. Moreover, SEN
and the SCE measure the spectral power distribution and the
center of " gravity " of the spectrum, respectively. These features
are widely used in speech recognition and perform well
in classification. In addition, a signal peak appears when the
whisker's tip first becomes stuck on a surface and then slips
off. Hence, it is reasonable to believe that the AIP reflects surface
texture information. Moreover, the AIP is a relatively
more stable feature than the peaks for each single signal point.
The four features can be calculated as follows:
ENGx t
SEN
=
=
SCE =
AIP =
/ 1
-
T
t=
/
/
/ ()i PP=2 ii 1
- -
N
N
,
(1)
where T is the length of the signal, x(t) is signal amplitude
at time t, sf
ber of peaks.
All the preceding extracted features are calculated and
scaled to [0,1] using the common normalization method to
compose a feature vector for subsequent processing. For the
4 × 40 training samples, all feature vectors and their labels,
which represent the texture type, were arranged to constitute
a feature matrix X and a corresponding label vector L,
as follows:
is the amplitude at frequency f in the discrete
Fourier transform of the signal, F is the Nyquist frequency,
Pi
is the time the ith peak occurs, and N is the total numContour
Reconstruction
To establish the contour reconstruction capability, we developed
a novel contour reconstruction algorithm for the whisker
sensor, based on its design principle and EBB theory.
The whisker sensor is designed to perceive contact by converting
deformation into pressure and measuring it by
piezoresistors. After perceiving contact, the relative position
of the contact point can be calculated by modeling the
DECEMBER 2022 * IEEE ROBOTICS & AUTOMATION MAGAZINE *
119
f
F
-/ ss=0 ff
log()
f
F
F
f
F
=
=
2 (),
0 fs$
0 sf
log()
,
R
X
Ll ll @.
=
=
6 12 160
g
T
S
S
S
S
AIP
SEN
SCE
ENG
1
1
1
1
AIP
SEN
SCE
ENG
2
2
2
2
g
AIP
SEN
SCE
ENG
160
160
160
160
V
X
W
W
W
W
,
(2)
In the training phase of the SVM model, because our aim
was to classify four textures, we used a one-versus-all strategy
to train four SVM binary classifiers. For each one, the optimal
hyperplane that separates the input data into two classes is
learned. The nth hyperplane can be defined as follows.
fx ay xx bnn nn
T
()=+ (3)
nn
.
Here, x denotes the 1 × 4 input feature vector, and a and b
represent the weight vector and bias, respectively. Substituting
feature matrix X and corresponding
label vector L
into (3), we note that for
each classifier, 160 sets of
data are used in training
to calculate weight an
and bias
b .n
After training,
a support vector
matrix Xn
S for the nth
classifier with a certain
weight vector an
s, bn
can be obtained.
In the classification
and bias
phase, when a test sample
vector x is input to
the system, four classifiers,
each with a certain support vector matrix
f
,
nk Xx
n ,,...,
= 12
=
argmax
12 4
n
S
.
(4)
Here, n represents the nth texture category, and kX xn
S
12
denotes the convolution operation. Thus, the texture of the
object scanned by the whisker can be determined using this
SVM model. The classifier with the highest convolution
score indicates the texture. In addition, the classification system
will output " none " when the scores for all four classifiers
are below 0.3, indicating that the test sample is not within the
scope of recognition.
The whisker sensor is
designed to perceive
contact by converting
deformation into pressure
and measuring it by
piezoresistors.
X ,n
S
perform
convolution operations with the sample vector and determine
the score for each category. Then, a majority vote
strategy is used to classify the test sample, as follows:
IEEE Robotics & Automation Magazine - December 2022

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2022
