In many pattern recognition and machine learning applications, researchers have focused on finding the best paradigm for the problem at hand: the one that guarantees the best performance according to whichever metric is considered appropriate to evaluate the paradigm. However, in many applications, there are innumerous other factors that should be considered, but are frequently overlooked. When it comes to Assistive Technology, overlooking other factors can exacerbate problems related to the extension of the disability. For example, if the user of the assistive technology requires assistance breathing, a technology based on sip-and-puff will not be of any help. Similarly,assistive technologies that require movement of the head, use of joystick, and many other motor skills become moot, even if the paradigm used -- e.g. sEMG on the neck or arm -- scores highly in, say, accuracy, recall, etc. In our experience, one muscle group that is rarely affected are facial muscles, in particular the movement of the eye-brows. So, in this research, we propose a comprehensive pipeline for eyebrow gesture classification, evaluating six distinct gestures using three different sensors: Surface Electromyography (sEMG), Inertial Measurement Unit (IMU), and video. The pipeline encompasses data collection from nine participants, data processing, and classification with a Multilayer Perceptron (MLP) and Resnet18 architecture. For the MLP, the mean accuracy of the network across all subjects exceeds 75 percent, with a maximum of 88 percent achieved for one participant. While Resnet18 the mean accuracy goes from 44 percent up to 85 percent for one participant. These results highlight the potential of eyebrow gestures as a powerful alternative to single-trigger events using a MLP as the selected classifier, offering a wide range of options in regards of sensors and gestures that can be tailored to the users' needs and capabilities.
Read more