Fast and Compact Ultrasound Features for Real-Time Hand Intent Detection in Human-Robot Interfaces

Kurzfassung

Reliable and low-latency biosignal-based intent detection is essential for seamless human-robot interaction in rehabilitation and assistive technologies. Fingertip force estimation provides a challenging instance of this problem, requiring methods that can capture subtle forearm muscle activity in real time. While ultrasound imaging has emerged as a promising modality, existing feature extraction methods often produce high-dimensional vectors or rely on computationally expensive operations, limiting their applicability in resourceconstrained systems. This paper introduces a novel feature extraction method based on efficient horizontal summation of B-mode ultrasound frames. The approach requires simple arithmetic operations, producing compact and discriminative features that preserve anatomical information while minimizing redundancy. Benchmarking against three established baselines-Histogram of Oriented Gradients, ROI-Gradient, and Image Moments-on a dataset of eight participants shows that the proposed method predicts fingertip forces with state-of-the-art accuracy (nRMSE=6.16%,R2=0.847), while reducing feature dimensionality by a factor of five and achieving more than four times faster extraction. A calibration analysis explores the reduction of required training data acquisition while maintaining high accuracy. These results highlight the method's potential for real-time deployment in wearable ultrasound systems, with a wide range of applications of human-robot interfaces.