Evaluation of Algorithms for Chew Event Detection

Analyzing food intake behavior is necessary to prevent obesity and overweight. Detecting and counting chewing strokes is an elementary part of this analysis. In our project, sounds of food intake were recorded using a microphone in the outer ear canal. The records contained sounds of 51 healthy subjects chewing 8 types of food. We evaluated seven different algorithms to detect chew events in sound records. Results of the automated detection were compared to manual annotations. Best performances (precision and recall over 76 %) were achieved by detecting chew events in six different frequency bands and fusing these results. With this method for counting the number of chews, an important step towards the estimation of bite weight has been done.