Hand Gesture Recognition via Wearable Ultra-Low Power Ultrasound and Gradient-Boosted Tree Classifiers

Wearable ultrasound (US) is becoming more popular for complementing surface electromyography in the hand gesture recognition (HGR) task. In fact, US allows collecting data from deep musculoskeletal structures with high spatiotemporal resolutions and high signal-to-noise ratio. However, existing wearable US solutions for HGR are not sufficiently low-power for guaranteeing continuous, long-term operation, and they typically rely on data processing and classification approaches not suited for edge computing. In this paper, we present the first armband for hand gesture recognition based on a truly wearable (12 cm
²
, 13 g), ultra-low power (16 mW) ultrasound probe, complemented by a lightweight classification approach based on XGBoost gradient-boosted tree classifier. We demonstrate an average cross-validated classification accuracy of 97% on four different gestures, while achieving low inter-session variability (standard deviation as low as 3%) in the scenario of armband removal and repositioning across experiments. Furthermore, thanks to its low complexity and memory usage (10 KB), the classifier can be executed in real-time on a low-power resource-constrained embedded platform. The system consumes only 16 mW and enables multi-day operation with a 320 mAh battery.