Proximity-based Eating Event Detection in Smart Eyeglasses with Expert and Data Models

Abstract

We compare performances of an expert model-based approach and a data-based baseline for eating event detection using proximity sensor data of smart eyeglasses. Proximity sensors in smart eyeglasses can provide dynamic distance estimates of cyclic temporalis muscle contraction during chewing without skin contact. Our expert model is based on proximity signal preprocessing and two-threshold grid search. In contrast, baseline data models were based on One-class Support-Vector-Machines. We evaluate both models with in-lab and free-living data from 15 participants. Free-living data were obtained across one day of wearing smart eyeglasses with temple-integrated proximity sensors in unconstrained settings. Overall, the retrieval performance F1 score of the two-threshold-based algorithm for free-living data ranged between 0.6 and 0.7, and outperformed all tested SVM model configurations. While SVM models achieved maximum recall, precision was often below 0.5. We report head-side specific performances for a bilateral arrangement of the proximity sensors and detail performance characteristics in model parameter sweeps. We conclude that eating detection using proximity sensors in smart eyeglasses is a promising approach for unconstrained automated dietary monitoring. Nevertheless, further investigations are needed to deal with the proximity signal characteristics in everyday life monitoring.