Acoustic Sensing for Fitness Activities Recognition: A Deep Learning Approach

Currently, the demand for home fitness has increased as more people choose to work remotely. Acoustic sensing technology has gained popularity due to its practicality on fitness activity monitoring. Due to the fact that sound waves do not propagate uniformly in all directions, identifying fitness activities can be challenging when sound waves do not directly face the device. To address it, we propose a method that models the reflected signal based on the sound field characteristics of a piston sound source and introduces a novel signal processing technique to reduce angle correlation. Then we combine deep learning methods for fitness activity recognition. Experimental results demonstrate improved recognition accuracy for fitness activities performed at different angles around the device.