IEEE Robotics & Automation Magazine - December 2017 - 109

Addition of EMG Features
The accuracies after adding EMG sensor data for the four
groups of motion classes did not increase considerably, as
shown in Figure 9. The average increase for the four groups
was roughly 0.83 ± 0.29%. Using the Student's t-test with a
p-value of 0.05, significant changes in classification accuracies
after the addition of EMG cannot be proven.
Discussion
FSR Placement
As presented in the "Results" section, configuration B provided a better performance (9.7% increase in mean classification accuracy) in six classes of motion as compared to
configuration A. However, classification accuracies
obtained from the two configurations were not considerably different for three classes. This shows that configuration B was able to capture the muscle movements
responsible for three additional grips, tripod, key, and finger point, better than configuration A. This is likely because
the change in the placement of the bands resulted in more
sensors being located over the bellies of the muscles that
were activated in finger movements rather than in grosshand movements. However, configuration A performed

LDA Classification Accuracy

FSR Placement
Neither of the configurations performed better for all groups
of motion classes. The configurations performed similarly for
the groups containing three grips and ten classes of motion.
Configuration B resulted in a better performance (9.7%
increase in mean classification accuracy) in six motion classes
compared to configuration A, while configuration A performed better (5.1% increase in mean classification accuracy)
in eight motion classes. The results obtained by the two configurations are shown in Figure 8.

better (a 5.1% increase in mean classification accuracy) in
eight motion classes. This is because the last four motion
classes in this group could be significantly distinguished
through more gross-hand movements.
For ten motion classes, the performance was almost
similar for the two configurations. This is because the
ten classes included both motion classes that involved
gross-hand movements and fine-finger movements. The
slightly better performance of configuration B for three
and ten motion classes is believed to be due to the placement of sensor strip 5, which can make different motion
classes more differentiable during vertical dynamic
movement. (See Table 2 for a summary of these classification accuracies.)
In a study [11], Radmand et. al. reported a machinelearning accuracy of 89% using high-density FMG and a
dynamic protocol similar to the one used in this study, with
motion classes similar to the motions in group number 3
in Table 1. The aforementioned study used ten healthy

FSR Placement

100
90
80
70
60
50
40
30
20
10
0
Three
Classes

Six
Classes

Eight
Classes

Ten
Classes

Number of Classes
FSR Placement A

FSR Placement B

Figure 8. The classification accuracies obtained for the sensor
placement investigation of three grips, six grips, eight motion
classes, and ten motion classes.

LDA Classification Accuracy

k = 30 was more generalized as compared to the one created
with k = 1, according to the bias-variance tradeoff [18]. At
this value of k, test accuracies were within the range of
68-72% for ten classes of motion.
Multiple kernels were tested for SVM to find the most
suitable one for this experiment's data, and the Gaussian kernel produced the best results. Gaussian SVM's hyperparameters were optimized for each case using cross validation, with
20% of the training data used for validation and 80% for
training the model. As previously mentioned, LDA and SVM
accuracies were obtained using the LOOCV method by leaving out each repetition (20% of all data in each experiment)
and averaging the results for five repetitions. These accuracies
are shown in Figure 7. Using the Student's t-test with a p-value
of 0.05, these algorithms' results did not prove to be significantly different. To compare our results with a study [11]
focused on dynamic protocols, classes of motion, and HMI
that are similar to the ones used in this study, the rest of the
offline data analysis described in this article used the LDA
classification algorithm.

100
90
80
70
60
50
40
30
20
10
0

Addition of EMG

Three
Classes

Six
Classes

Eight
Classes

Ten
Classes

Number of Classes
Without EMG

With EMG

Figure 9. The classification accuracies obtained with the addition
of EMG for three grips, six grips, eight motion classes, and ten
motion classes.

DECEMBER 2017

IEEE ROBOTICS & AUTOMATION MAGAZINE

109

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