ProxiFit

Abstract

Although many works bring exercise monitoring to smartphone and smartwatch, inertial sensors used in such systems require device to be in motion to detect exercises. We introduce ProxiFit, a highly practical on-device exercise monitoring system capable of classifying and counting exercises even if the device stays still. Utilizing novel proximity sensing of natural magnetism in exercise equipment, ProxiFit brings (1) a new category of exercise not involving device motion such as lower-body machine exercise, and (2) a new off-body exercise monitoring mode where a smartphone can be conveniently viewed in front of the user during workouts. ProxiFit addresses common issues of faint magnetic sensing by choosing appropriate preprocessing, negating adversarial motion artifacts, and designing a lightweight yet noise-tolerant classifier. Also, application-specific challenges such as a wide variety of equipment and the impracticality of obtaining large datasets are overcome by devising a unique yet challenging training policy. We evaluate ProxiFit on up to 10 weight machines (5 lower- and 5 upper-body) and 4 free-weight exercises, on both wearable and signage mode, with 19 users, at 3 gyms, over 14 months, and verify robustness against user and weather variations, spatial and rotational device location deviations, and neighboring machine interference.