A study of time-based features and regularity of manipulation to improve the detection of eating activity periods during free living

This thesis considers the problem of detecting when people eat by tracking their wrist motion. The goal of this work is to automatically detect the start and stop times of these eating activities. It builds upon previous work done by our research group [12] which developed an algorithm for automatically detecting peaks in activities associated with food preparation and clean-up. This peak detector is then used for segmenting data. These segments are then classified as eating or non-eating activities using a naive Bayes classifier based on probabilities obtained from computing different features in each segment. The original work introduced 4 features, all of them based on sensor readings. In this thesis we introduce a set of 3 new features to improve the detection of eating and non-eating activity periods: regularity of manipulation, time since last eating activity and cumulative eating time. We discuss the main concepts behind them, introducing the idea of time-based features. We then we test our new features under the framework developed by our group. Detection including regularity of manipulation, in combination with the original 4 features, achieved an overall accuracy of 79%. The accuracy obtained including time since last eating activity reached 69% and cumulative eating time, an overall accuracy of 64%. Finally, we compare these results to the original work and later discuss and characterize results based on our findings.