Using myoelectric signals for gesture detection: a feasibility study

Abstract
The propose of this study was to assess the feasibility of using myoelectric
signals acquired using an off the shelf device, the Myo armband from
Thalmic Lab.
Background:
With the technological advances in sensing human motion, and its potential
to drive and control mechanical interfaces remotely, a multitude of input
mechanisms are used to link actions between the human and the robot. In
this study we explored the feasibility of using human arm’s myoelectric signals
with the aim of identifying a number of gestures automatically.
Material and methods:
Participants (n = 26) took part in a study with the aim to assess the gesture
detection accuracy using myoelectric signals. The Myo armband was
used worn on the forearm. The session was divided into three phases, familiarisation:
where participant learned how to use the armband, training:
when participants reproduced a number of requested gestures to train our
machine learning algorithm and recognition: when gestures presented on
screen where reproduced by participants, and simultaneously recognised using
the machine learning routines.
Results:
One participant did not complete the study due to technical errors during the
session. The remaining (n = 25) participants completed the study allowing
to calculate individual accuracy for grasp detection using this medium.
Our overall accuracy was 65.06%, with the cylindrical grasp achieving the highest accuracy of around 7.20% and the tripod grasp achieving lowest
recognition accuracy of 60.15%.
Discussions:
The recognition accuracy for the grasp performed is significantly lower compared
to our earlier work where a mechatronic device was used. This could
be due to the choice of grasps for this study, as it is not ideal to the placement
of the armband. While tripod, cylindrical and lateral grasps have different
finger and wrist articulations, their demand on supporting forearm muscles
(mainly biceps and triceps) is less definite and therefore their myoelectric
signals are less distinct. Furthermore, drop in accuracy could be caused
by the fact that human muscles and consequently the myoelectric signals
are substantially variable over time. Muscles change their relative intensity
based on the speed of the produced gesture. In our earlier study, the gesture
production speed was damped by the worn orthosis, leading to normalising
the speed of gestures. This is while in our current study, hand motion is not
restricted. Despite these, the recognition accuracy is still significant.
Future work:
There are remaining questions related to the feasibility of using myoelectric
signals as an input to a remote controlled robot in a factory floor as it is
anticipated that such a system would enhance control and efficiency in production
processes. These questions therefore require further investigations
regarding usability of the armband in its intended context, to ensure users
are able to effectively control and manipulate the robot using the myoelectric
system and enjoy a positive user experience. Future studies will focus
on the choice of gestures, so that they are distinct and better identifiable,
but also on other key human factors and system design features that will
enhance performance, in compliance with relevant standards such as ISO
9241-210:2010 (standards for human-system interaction ergonomic design
principles) . Furthermore, aspects of whether a machine learning algorithm
should use individually learned events in order to recognise an individual’s
gestures, or if it is possible to use normative representation of a substantial
set of learnt events, to achieve higher accuracy remains an
interesting area for our future work.