IEEE Robotics & Automation Magazine - December 2017 - 107

FSR Placement
One factor affecting system performance was the position of
FSR sensors inside the inner socket. Two different configurations, A and B, were investigated and compared [as shown in
Figure 6(a) and (b)]. In configuration A, FSR strips 1 and 3
were placed on the extensor muscles, and sensor strips 2 and 4
were placed over the flexor muscles. These four sensor strips
were placed so that they would cover the muscle bellies while
not covering the EMG sensor cutouts. FSR strip 5 was placed
circumferentially over the muscle bellies. In configuration B,
FSR strips 1 and 2 were positioned just distal to EMG sensors
to capture the maximum amount of contraction of the extensor and the flexor muscles. In this configuration, sensor strip 3
was placed where sensor strip 1 was located in configuration
A, and sensor strip 4 remained in the same location in both
configurations. FSR strip 5 was placed on the anterior side of
the forearm.
The number of active FSR sensors in the two configurations differed. In configuration A, 58 active sensors were
embedded in the inner socket; however, to avoid the cutouts
provided for EMG sensors, these strips could not be placed
directly over the muscle bellies, with the most palpable deformation during contractions controlling gross-hand movement. In configuration B, two of the FSR strips were placed
directly on the muscle bellies under the EMG sensors; however, this resulted in a lower number of active sensors (37). The
other difference in these two configurations was the fifth

sensor strip. In the configuration A placement, it covered the
muscle bellies circumferentially, while, in configuration B, the
fifth FSR strip was placed so that it could catch the effect of
the weight of the socket during vertical dynamic movement.
Data Analysis
Data analysis was performed on the four classes of motion
groups represented in Table 1. The analysis was performed offline. To examine the performance of the system
in different experiments, machine-learning accuracies
were compared. Numerous machine-learning algorithms
have been used in similar studies to classify hand gestures.
In this study, we examined three algorithms that are
among the most popular reported in literature: k-nearest
neighbors (kNN), support vector machine (SVM), and
linear discriminant analysis (LDA) [17]. Classification
accuracies were obtained through the leave-one-out cross
validation (LOOCV) method. The reported results are the
mean and standard deviation of the accuracies obtained
by leaving out each of the five repetitions. To determine
the significance of the difference of the results obtained by
these algorithms, the Student's t-test with a p-value of 0.05
was used.
Addition of EMG Features
One of the factors examined in this study is the effect of
including EMG signals as additional features. For this

FSR Strip 5

FSR Strip 4

EMG
Cutout

FSR Strip 1

FSR Strip 5

FSR Strip 2
EMG
Cutout
FSR Strip 3

FSR Strip 1

FSR Strip 4
FSR Strip 3

FSR Strip 2

(a)

(b)

Figure 6. The FSR placement: (a) configuration A and (b) configuration B.

DECEMBER 2017

IEEE ROBOTICS & AUTOMATION MAGAZINE

107

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2017