Hand Signs Recognition by Deep Muscle Impedimetric Measurements

Abstract

This study investigates the potential of impedimetric measurements providing deep muscle information for recognizing 36 American Sign Language (ASL) hand signs. Two measurement methods are considered together for the first time: electrical impedance myography (EIM) and electrical impedance tomography (EIT). EIM was measured along the anterior forearm, while 8-electrode EIT was recorded around the forearm below the elbow. Data were acquired from three volunteers, with each hand sign performed ten times. A correlation analysis was conducted to identify the relevant EIM frequencies to distinguish between hand signs. Among all evaluated algorithms, the random forest classifier achieves the highest classification performance. Classification based on the resistance and reactance at the selected EIM frequencies achieved $61.54~{\pm }~0.85$ %, while classification based on EIT boundary voltages achieved 91.04% ${\pm }~0.46$ %. Combining the results from both classifiers into an EIM-EIT hybrid classifier improved the accuracy to $92.57~{\pm }~0.41$ %, effectively reducing ambiguities between similar hand signs. Achieved results considerably outperform state-of-the-art works, which typically classify fewer hand signs or achieve lower accuracy.