Classification of Individual Finger Flexions Using Ultrasound Radiofrequency Signals

The objective of this thesis is to develop a wearable human machine interface (HMI) to classify individual finger flexions using ultrasound. To study the effect that lateral resolution has on the finger classification performance, 127 ultrasound radiofrequency (RF) signals are acquired from a 40 mm width probe. The acquired signals were reconstructed to show an estimated number of 4 - 8 reconstructed RF signals can maintain high accuracies to that of full resolution. The pattern classification pipeline of this study, using a computationally efficient spatial feature extraction method novel ultrasound-based studies, is verified to make accurate finger predictions every 100 ms in time throughout the full finger flexion recordings. Finally, using three 280µm thin wearable ultrasonic sensors attached to the forearm achieved a classification accuracy of 97.5 ± 2.9%. The results of this thesis provide the guidelines to design for wearable HMI applications.