Physical Activity Recognition and Identification System

Abstract

Background: It is well-established that physical activity is beneficial to health. It is less known how the characteristics of physical activity impact health independently of total amount. This is due to the inability to measure these characteristics in an objective way that can be applied to large population groups. Accelerometry allows for objective monitoring of physical activity but is currently unable to identify type of physical activity accurately. Methods: This thesis details the creation of an activity classifier that can identify type from accelerometer data. The current research in activity classification was reviewed and methodological challenges were identified. The main challenge was the inability of classifiers to generalize to unseen data. Creating methods to mitigate this lack of generalisation represents the bulk of this thesis. Using the review, a classification pipeline was synthesised, representing the sequence of steps that all activity classifiers use. 1. Determination of device location and setting (Chapter 4) 2. Pre-processing (Chapter 5) 3. Segmenting into windows (Chapters 6) 4. Extracting features (Chapters 7,8) 5. Creating the classifier (Chapter 9) 6. Post-processing (Chapter 5) For each of these steps, methods were created and tested that allowed for a high level of generalisability without sacrificing overall performance. Results: The work in this thesis results in an activity classifier that had a good ability to generalize to unseen data. The classifier achieved an F1-score of 0.916 and 0.826 on data similar to its training data, which is statistically equivalent to the performance of current state of the art models (0.898, 0.765). On data dissimilar to its training data, the classifier achieved a significantly higher performance than current state of the art methods (0.759, 0.897 versus 0.352, 0.415). This shows that the classifier created in this work has a significantly greater ability to generalise to unseen data than current methods. Conclusion: This thesis details the creation of an activity classifier that allows for an improved ability to generalize to unseen data, thus allowing for identification of type from acceleration data. This should allow for more detailed investigation into the specific health effects of type in large population studies utilising accelerometers.