Monday, August 8, 2016 at 2:00 PM in Rice 504
Advisor: Kamin Whitehouse
Committee Members: Jack Stankovic (Committee Chair), Hongning Wang, Quanquan Gu (Systems Engineering), David Culler, (UC Berkeley).
Toward Robust and Accurate Human Activity Recognition using Wearable Sensors
Ubiquitous and immersive sensing equipment and devices, e.g., the Internet of Things, are generating an explosive amount of data. To extract meaningful and actionable information out of these data inevitably requires the metadata of the generated data. However, the majority of data remain unlabeled and the generation of metadata still involves labor intensive efforts, thus fundamentally unscalable. As a solution, we propose a framework for inferring the contextual information embedded in the sensor time series, e.g., what they measure, where they locate, how they relate to each other, etc. We take a representative of the metadata inference problem, particularly, for commercial buildings, and demonstrate first steps towards a metadata inference solution that requires minimal human intervention. At core of our solution lie a suite of techniques that exploit both the textual and time series data of sensors in buildings. We have explored a few approaches to inferring the type and location information that show promise. Building upon the early results, we will next focus on inferring the relationship between points. This proposal provides an overview of our solution, along with some preliminary results and key challenges, proposed research and an evaluation plan.