Abstract
Smart home sensor systems that infer residents' activities enable a number of exciting medical monitoring and energy conservation applications. Existing home activity recognition systems are invasive, since they require significant manual effort from the end user in installing or training the system, inconvenience the residents by requiring them to constantly wear tags, or require invasive and expensive sensors. Our hypothesis is that by effectively using data fusion techniques, leveraging the existing smart meter infrastructure, and using only weak biometric sensing, we can build convenient, accurate home activity recognition solutions.
The key home activity recognition challenges addressed in this dissertation include reducing configuration effort from the end user, reducing sensor installation effort, and identifying residents in multi-person homes without using invasive sensors. To reduce configuration effort, we develop an unsupervised activity recognition algorithm called AUTOLABEL that leverages data fusion and cross-home activity models to accurately recognize resident activities without user training. To eliminate many direct sensors in the home for activity recognition, we develop effective Bayesian data fusion techniques, which combine the existing smart meter infrastructure in homes, with a low cost, non-invasive sensor per room. Finally, we propose the use of resident height, which is a weak biometric, to identify residents for activity recognition in multi-person homes.
We evaluate our proposed activity recognition solutions through short term prototype sensor deployments in homes lasting from 7 to 10 days each. We show that our solutions satisfy the activity recognition needs of numerous smart home applications, such as remote medical monitoring, and fine-grained resource consumption monitoring of light and water fixtures in the home. Finally, we observe that our unsupervised activity recognition algorithm can be used in a wireless snoop attack on smart homes, to infer the residents' daily activities with high accuracy in spite of encrypted wireless transmissions. We propose and evaluate a suite of privacy solutions to mitigate the inference accuracy of such an attack without affecting the performance or functionality of the home activity recognition system.
