A study of the regularity of autocorrelation of manipulation of wrist motion during eating

This thesis considers the problem of detecting the eating activities (e.g. meals, snacks) of people by tracking their wrist motion. Our goal is to automatically detect the start and end times of eating activities during free-living. It builds upon previous work done by our research group [6] [7] [8] [21]. The detection is done by segmenting data into segments then classifying the segments as EA (eating activity) and NonEA (non-eating activity) using a Bayesian classifier. Previous features used in the classifier developed by our group included the sum of acceleration, the amount of manipulation, the amount of wrist roll and the regularity of wrist roll [6] [7] [8]. Additional features studied include a frequency analysis of manipulation, the time since the last EA, and the cumulative time spent eating in a day [21]. In this thesis we study two new features: the autocorrelation of the manipulation and the off-line analysis of the time since the last EA. The autocorrelation feature enables the study of patterns of manipulation that may not be precisely regular, and facilitates frequency analysis through the transform of the manipulation signal into something more sinusoidal. The off-line analysis of the time since the last EA allows for the calculation of every possible combination of segment classifications throughout the day, so that an early incorrect classification of a segment as an EA does not always inadvertently affect the classification of subsequent segments. This thesis further discusses these concepts and then tests them under the framework developed by our group. The