University of Twente Student Theses

Login

A wearable sensor system for eating event recognition using accelerometer, gyroscope, piezoelectric and lung volume sensors

Mevissen, Sigert (2021) A wearable sensor system for eating event recognition using accelerometer, gyroscope, piezoelectric and lung volume sensors.

[img] PDF
10MB
Abstract:Overweight and obesity are a large and increasing problem worldwide and have been associated with a range of diseases. People are advised to exercise more, eat less but more regularly and avoid unhealthy food to lose weight. The most important principle is to burn more calories than you consume. Food consumption can be registered using a food diary, but entries are often inconsistent by over- or underestimating the amount of food consumed or by forgetting to register consuming food at all. This leads to the introduction of automatic dietary monitoring, which aims to objectively measure on the consumption of food by determining the timing of food consumption and the quantity and type of food consumed. A sensing system consisting of a smartwatch, a piezoelectric sensor and a respiratory inductance plethysmography sensor has been introduced to detect eating events. The smartwatch worn on the wrist consists of a gyroscope and an accelerometer in three axes to detect eating gestures. A piezoelectric sensor worn on the jaw is used to recognise chewing food. The respiratory inductance plethysmography sensor consists of two bands that measure lung volume change to detect swallowing food. The sensor data of different sensors is combined to find out how much they complement each other. An experiment was conducted in which six participants are asked to eat a croissant and a bowl of yogurt with pieces of apple. The sensor data is filtered, transformed and then split into windows. Distinguishing characteristics of the windows are captured by creating features. These features are fed into the classification algorithms to classify the windows as eating events or non-eating events. Different machine learning configurations are implemented to determine the optimal feature vector length, sensor combination and to test the generalisability of the model. In addition, a finite-state machine is implemented to capture the sequential nature of eating by filtering false positive eating event classifications. The highest F$_1$-score found using only machine learning algorithms is 0.82 for the classification of eating gestures, 0.94 for chewing food and 0.53 for swallowing food. The finite-state machine did not improve the results. The classification of eating events and chewing food achieved high scores, and the smartwatch and piezoelectric sensor are therefore effective in the detection of eating events. The current configuration of the respiratory inductance plethysmography sensor is ineffective in the detection of swallowing food. Future research can focus on how the detection of eating events translates to the detection of eating by measuring in less-controlled measurement conditions.
Item Type:Essay (Master)
Faculty:TNW: Science and Technology
Subject:02 science and culture in general
Programme:Biomedical Engineering MSc (66226)
Link to this item:https://purl.utwente.nl/essays/88431
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page