Recognizing and Discovering Activities of Daily Living in Smart Environments

Identifying human activities is a key task for the development of advanced and effective ubiquitous applications in fields like Ambient Assisted Living. Depending on the availability of labeled data, recognition methods can be categorized as either supervised or unsupervised. Designing a comprehensive activity recognition system that works on a real-world setting is extremely challenging because of the difficulty for computers to process the complex nature of the human behaviors. In the first part of this thesis we present a novel supervised approach to improve the activity recognition performance based on sequential pattern mining. The method searches for patterns characterizing time segments during which the same activity is performed. A probabilistic model is learned to represent the distribution of pattern matches along sequences, trying to maximize the coverage of an activity segment by a pattern match. The model is integrated in a segmental labeling algorithm and applied to novel sequences. Experimental evaluations show that the pattern-based segmental labeling algorithm allows improving results over sequential and segmental labeling algorithms in most of the cases. An analysis of the discovered patterns highlights non-trivial interactions spanning over a significant time horizon. In addition, we show that pattern usage allows incorporating long-range dependencies between distant time instants without incurring in substantial increase in computational complexity of inference. In the second part of the thesis we propose an unsupervised activity discovery framework that aims at identifying activities within data streams in the absence of data annotation. The process starts with dividing the full sensor stream into segments by identifying differences in sensor activations characterizing potential activity changes. Then, extracted segments are clustered in order to find groups of similar segments each representing a candidate activity. Lastly, parameters of a sequential labeling algorithm are estimated using segment clusters found in the previous step and the learned model is used to smooth the initial segmentation. We present experimental evaluation for two real world datasets. The results obtained show that our segmentation approaches perform almost as good as the true segmentation and that activities are discovered with a high accuracy in most of the cases. We demonstrate the effectiveness of our model by comparing it with a technique using substantial domain knowledge. Our ongoing work is presented at the end of the section, in which we combine pattern-based method introduced in the first part of the thesis with the activity discovery framework. The results of the preliminary experiments indicate that the combined method is better in discovering similar activities than the base framework.