Discovering Frequent ADL Patterns From Wearable Accelerometers

The discovering and understanding of human physical activity has long been the research interest of wearable computing and ubiquitous computing. The generally accepted approach is to use some offline machine learning techniques to map the human physical data to pre-defined labels. However, the personalized modeling raises new challenges. One of the challenges is that the user’s activity may not occur in the pre-defined label list. This issue becomes more practical when it comes to the understanding of activities of daily living(ADLs), since people perform different ADLs according to their professions. Therefore, it requires the system to automatically discover these out-of-vocabulary activities, especially those who occupy most of the time.In this thesis, we propose an unsupervised method to discover frequent ADL patterns from wearable acceleration data streams. This approach discovers frequently-occurred ADL patterns and learns a classifier for each discovered pattern. It differs from existing motion discovery approaches in three ways. First, the number of discovered patterns is dynamically computed, which is adaptive to the data itself. Second, an anonymous classifier is built for every discovered pattern, which could be reused to predict activities after naming. Third, this approach can keep digging new patterns when new data stream comes.The main research and contributions in this thesis include the following aspects:(1) Designed an ADL pattern discovering and clustering approach. In this approach, the number of discovered pattern is dynamically computed. And it can keep digging for new patterns of activity when new acceleration stream comes.(2) Proposed a segmentation method based on the steadiness of topic distribution. This method effectively discovers potential activity patterns by slicing continuous motion samples with the same pattern into one segment. We perform a ‘one-steady-topic-distribution-to-one-pattern’ segmentation instead of ‘one-topic-to-one-pattern’ method, which effectively separates the number of discovered patterns from the number of topics.(3) Proposed a training data sampling method. This method automatically generates the training samples and their labels from segmentation results. The generated training samples are of reasonable numbers and diversity.(4) Proposed an iterative discovery method with reusable classifiers, which grants our method the ability of continuous mining.(5) Evaluated our approach on Ubicomp08 and PAMAP2 dataset. The result indicates that our approach has good ability in ADL discovering. It discovers 13 ADL patterns from Ubicomp08 dataset. The discovered patterns cover 86% raw data, with 82% sensitivity and 80% clustering accuracy. In the experiment on PAMAP2 dataset, altogether 17 patterns are discovered and 67% data are covered. The approach reaches 92% sensitivity and 91% clustering accuracy. Also, the learned classifiers achieve 71% sensitivity and 84% accuracy in predicting activities on other days of data in Ubicomp08 dataset.