Towards Whole Hand and Wrist Kinematic Tracking with a Wearable A-Mode Ultrasound Probe

This paper introduces a robust, wearable framework for tracking hand and wrist kinematics using A-mode ultrasound (US). While previous research has struggled to balance high-dimensional tracking (23 degrees of freedom) with the power and memory constraints of wearable devices, this work demonstrates a successful end-to-end embedded implementation. The authors utilize the WULPUS platform to perform data acquisition, neural network inference, and result streaming entirely on-device.

The researchers employed a compact convolutional neural network (11,285 parameters) to regress 23 joint angles from ultrasound data. To address the common problem of accuracy degradation when a wearable sensor is removed and repositioned, they implemented a user-centric incremental fine-tuning strategy. This allows the model to adapt to new sensor placements using only about 5 minutes of calibration data. The entire pipeline was deployed on an nRF52832 microcontroller, utilizing TensorFlow Lite Micro and 8-bit integer quantization to fit within strict memory and power budgets.

The system achieves a 29.1 ms inference latency and consumes less than 35 mW, enabling over 36 hours of continuous operation on a standard battery. The incremental fine-tuning approach reduced the mean absolute error (MAE) by over 17% compared to non-incremental baselines. Furthermore, by performing inference on-device and transmitting only the resulting joint angles, the system achieves an 88% reduction in wireless bandwidth usage compared to streaming raw ultrasound data.

This work provides a viable path for integrating ultrasound-based sensing into commercial prosthetics and human-machine interfaces. By demonstrating that high-dimensional kinematic tracking can be performed with low power and high robustness to sensor repositioning, the authors address the primary barriers to the practical, long-term deployment of wearable ultrasound technology.