Probabilistic activity recognition from low-level sensors

Activity recognition is gaining increasing interest in the artificial intelligence (AI) and ubiquitous computing communities due to the dramatically advancing sensor technology. In this dissertation, we address the problem of probabilistic activity recognition from low-level sensor data. The novelty of our work can be seen from two fronts. First, in the pervasive computing literature, an important focus has been to determine a user's context from streams of sensor data. Despite the large amount of previous work done on computing the locations of users, there has been a lack of study on the problem of high-level activity recognition. Second, in the AI area, recognizing complex high-level behavior has traditionally been the focus of plan recognition. However, most of the work has been restricted to high-level inferences in a logical framework, and the challenge of dealing with low-level sensor modeling has so far not been adequately addressed. In this dissertation, we propose several novel probabilistic algorithms for activity recognition from low-level sensors. Firstly, we present a novel clustering algorithm to automatically group user traces into a set of clusters, each corresponding to a typical class of user activity patterns. Secondly, we propose a hierarchical activity recognition model, in which high-level inference about activities is enabled via a location-based sensor model at the low level. Thirdly, to reduce the calibration effort for activity recognition and to increase robustness in recognition quality, we design a segmentation-based activity recognition model, in which activities can be directly recognized from sequences of discovered motion patterns. Finally, we also propose a novel approach to detecting users' abnormal activities from sensor data. We demonstrate the effectiveness of our proposed algorithms using the data collected in real wireless environments.