A Microcontroller-Based Platform for Cognitive Tracking of Sensorimotor Training

Abstract

A new generation of fitness trackers is pervasively invading different aspects of our life, taking profit from wireless technology, embedded sensors, and increasingly accurate AI-based data analysis. The most crucial aspects concerning the design of these systems include energy efficiency and accuracy. In this paper, we propose a system relying on two microcontroller-based sensor nodes to track the physical activity during sensorimotor training, a type of exercise that challenges the user’s balance skill, which has been proven to be very effective in improving performance, preventing injuries and recovering from them. One of the sensor nodes is integrated into a custom wobble-board and the second is wearable by the user. The nodes are adaptable to be set in different operating modes, depending on the use case needs, enabling different steps of near-sensor pre-processing. The most power-efficient operating mode executes a CNN-based analysis directly on the microcontroller, to recognize physical exercises. The algorithm provides an accuracy of respectively 99.4% and 97.6% on the two nodes. In-place execution of the CNN saves up to 65% power consumption with respect to the transmission of raw data for on-cloud analysis.

Keywords

• Adaptive system
• fitness activity tracking
• sensorimotor training
• low power electronics
• neural network
• remote sensing