Automatic Segmentation Method of Bone Conduction Sound for Eating Activity Detailed Detection

Many studies have shown that a low chewing rate during meals leads to obesity. Also, presenting the chewing amount in real-time to eaters prevents them from eating too fast and improves their eating awareness. This study aims to provide a system that improves the awareness of good eating habits by real-time quantification of eating activity in a natural meal environment. The concrete object of this research is to develop a method to segment automatically eating detailed activities accurately. We collected meal sound data from 14 subjects in a natural meal environment using a bone conduction microphone. Several people did data labeling, keeping only similar labels to make a robust dataset. This paper proposed a method that automatically segments the bone conduction sound corresponding to the eating detailed activities. The evaluation of precision was 88.1%, and the average recall of each class was 70.5%.