Recognizing Eating Gestures Using Context Dependent Hidden Markov Models
Yiru Shen, Eric R. Muth, Adam W. Hoover · 2016
This paper considers the problems of recognizing eating gestures by tracking wrist motion. Hidden Markov models (HMMs) were developed to capture variations in motion patterns of subgroups of participants. Specifically, we examined if foreknowledge of the gender, age, and utensil used for eating could improve recognition accuracy. Improvement in accuracy was measured by comparing to a baseline HMM that was trained on all participants. Data was collected for 276 participants eating a single meal within a cafeteria setting. A total of 44,873 gestures were manually labeled using video synchronized with the wrist motion tracking device. Results show that gender HMMs performed slightly better than the baseline, indicating that there is not much difference in wrist motion patterns during eating between females and males. Age HMMs provided a 4.3% increase in accuracy and utensil HMMs provided a 6.2% increase inaccuracy. The results suggest that contextual variables can be used for improving gesture recognition.