Detecting food intake acoustic events in noisy recordings using template matching

Abstract

Automatic food intake monitoring using wearable systems is a promising research direction in the fight against obesity and eating disorders. Previous work has indicated progress towards automatic food intake monitoring using acoustic sensors for detecting periods of food intake, swallowing and chewing detection, and discriminating between solid and liquid food intake. However, little effort has been put towards acoustic detection in noisy recordings. In this paper, we explore detecting food intake events (particularly chew events) in the presence of restaurant background noise. Three templates were extracted from a clean signal to represent the beginning, middle and end phase of a chewing sequence. Then, each template was used with sliding window correlation to detect chew events in a noisy recording. The noisy signal was formed by instantaneous addition of a clean throat microphone recording during eating and restaurant noise recorded with the same throat microphone. Results show that the template created from the end phase of a chewing sequence outperformed templates from the beginning and middle phases for detecting chew events in a continuous clean and noisy test signal. An F1 score of 71.4% was achieved for detecting chews in very low signal-to-noise ratio of −10 dB.