Computer Science Colloquia
Friday, April 25, 2014
Advisor: John A. Stankovic
Attending Faculty: Kamin Whitehouse (Chair), Alfred Weaver, John Lach (minor representative), and Stephen Patek
9:00 AM Rice Hall, Rm. 242
PhD Dissertation Defense Presentation
Addressing Realisms in Activity Recognition for Smart Home Deployments
Research in wireless sensor networks has been very successful in creating test beds and short-term deployments for many application areas (e.g, home health care, saving energy in buildings, security systems) that depend on accurate activity recognition. The utility of these activity recognition systems often depends on recognizing anomalies from typical behaviors learned based on the activities. However, for many real home situations these activity recognition and anomaly detection solutions are not robust enough due to many realities.
In this dissertation, we have designed, implemented and evaluated a novel activity recognition system named AALO, a comprehensive anomaly detection system in daily activities named Holmes, and a novel ground truth collection system named Vocal-Diary that can be used to evaluate both AALO and Holmes. AALO is an active learning based activity recognition system that applies machine learning and data mining techniques to address some of the realities of deployments including difficulty in obtaining labeled ground truth for training, individuals performing overlapping activities, and generalizability in varying smart home environments. Holmes is a comprehensive anomaly detection system for daily in-home activities that learns normal variability in daily activities based on specific days of the week, combine activity instances of a day / multiple days together to find the features that best represent regularity, and detect temporal and causal correlations among multiple activities. In addition, Holmes uses semantic rules learned based on resident and expert feedback that explain specific variations in daily activities in specific scenarios. Finally, Vocal-Diary is a voice command based ground truth collection system where residents log activities by specific voice commands. To ensure robustness in the presence of different environmental noise in the home, Vocal-Diary integrates two-way acknowledgements and speaker recognition in the framework.
We evaluate the contributions with publicly available datasets and our own deployments. Results show that Holmes and Vocal-Diary performs better than state of the art systems and AALO performs as good as the state of the art supervised activity recognition systems without requiring the large amount of ground truth that they need.