Abstract:
Recognizing human behavior in an automated manner is essential in many ambient intelligence applications such as smart homes, health monitoring applications and emergency services. In order to make such long term health monitoring systems sustainable, we need smart environments in which the human activities are recognized automatically. In order to infer the human behavior, we can use machine learning methods on the data collected from the smart environments but those methods require annotated datasets to be trained on. Recording and annotating such datasets are costly since they require time and human e ort. Moreover, the complex nature of human activities makes it di cult to accurately model them. While hierarchical models can be a remedy for more accurate representation, nding suitable complexity levels is not a trivial task. Finally, when we deploy automatic human behavior monitoring systems on a world-wide scale, we need to ne tune the model behavior for each new house to accurately re ect the residents' behavior for that speci c house. Rather than annotating a dataset consisting of several weeks of data, an algorithm can be used to decide for which point in time it would be most informative to obtain annotation in order to minimize the need for annotation and maximize the usefulness of annotation. This thesis addresses the above mentioned issues by (i) collecting publicly available benchmark datasets, (ii) proposing a methodology for incorporating a hierarchy into the model that is tailored for various activities individually, (iii) improving the ways of evaluating di erent approaches and models considering the domain speci c needs, (iv) handling multi-resident environments in an unobtrusive manner and, (v) using active and semi-supervised learning techniques in order to reduce the annotation e ort in large scale deployments.
