Learning-Based Calibration Decision System for Bio-Inertial Motion Application

We developed a learning-based calibration algorithm for a vestibular prosthesis with the long-term goal of reproducing error-free vestibular system dynamic responses. Our approach uses an additional IMU to detect the head acceleration of a patient and to correct the corresponding drift in the vestibular prosthesis. The algorithm includes four major parts. First, we extract features from the shoe-mounted IMU to classify human activities through convolutional neural networks. Second, we fuse data from the head-mounted IMU (vestibular prosthesis). Third, we artificially create additional data samples from a small pool of training data for each classification class. Fourth, we use the classified activities to calibrate the reading from the head-mounted IMU. The results indicate that during daily routine activities the firing rate baseline of a vestibular prosthesis system without calibration fluctuates between 100 pulses/s to 150 pulses/s; in contrast, an appropriate calibration to human activity results in correction of 4 pulses/s in extreme cases, providing a stable baseline firing rate while the head is not moving. In this work, we specifically study the contribution of gyroscope scale factor on the drift of the vestibular prosthesis system and propose a corresponding calibration method.