A Sensory Glove With a Limited Number of Sensors for Recognition of the Finger Alphabet of Polish Sign Language

Existing data glove designs are fitted with multiple sensors. This is because existing designs often prioritize accuracy at the cost of ergonomics, accessibility, and affordability. This paper addresses this gap by building a model of a novel sensory glove with a reduced number of sensors and using machine learning paradigms in experiments to maximize gesture recognition performance. Usability tests of the data glove were carried out with 15 participants performing static and dynamic gestures in the alphabet of Polish Sign Language (APJM). The influence hierarchy of individual piezoelectric sensors was determined using a decision tree algorithm during previous stage of research, which achieved 94% accuracy with data from only three sensors. For dynamic gesture recognition, a neural network with convolutional layers and GRU recursive units achieved 99% accuracy using data from all sensors. The main contribution of the work is the identification of a universal set of three piezoelectric sensors and six inertial sensor axes for classification accuracy exceeding 98% for APJM letters. The results support the idea of simplified sensory glove designs with fewer than five sensors, promising improved ergonomics, reliability, and accessibility for people with disabilities.