The goal of this Faculty Early Career Development Program (CAREER) project is to enable "smart buildings" that can locate and identify specific individuals, and classify their activity, based only on the vibrations of the building structure caused by footsteps. Elder care facilities aim to maintain or improve the quality of life and independence of elders while reducing costs and capacity needs for care-professionals. One key to achieving this goal is to understand the activities of each occupant. Existing solutions to monitor occupants, such as vision, acoustic, motion, and force sensors and mobile devices, have strict installation requirements. These requirements lead to intrusive and dense deployment or require active user involvements. Instead, this project is built upon sensing the vibrations created by occupants' during their walking activity. Using building vibration to monitor occupants allows non-intrusive and scalable monitoring with inexpensive vibration sensors. More generally, this research will enable smart buildings to sense, track, and predict the status of occupants in a maintainable way using "structures as sensors" and thus enable future occupant-aware applications. Similarly, the technology can locate portions of a building with slippery or unsafe footing, or detect the presence of unauthorized people in restricted areas. By tracking first responders and locating imperiled civilians, such systems will also help dispatchers to mitigate emergencies. The project includes proof-of-concept deployments in three different elder care facilities. Targeted outreach activities will highlight the capabilities of this technology at an appropriate level of detail to appeal to female middle-school students.
This project uses structures themselves as activity sensors, by passively sensing footstep-induced floor vibrations, and employing advanced sparse-signal approximation in a Bayesian framework, to extract individual activity information. The specific research thrusts are: (1) extracting individual persons' footstep-induced floor vibration signal from a noisy signal mixture due to multiple human sources, by exploiting hierarchical wavelet decompositions and applying structured sparsity regularization; (2) localizing individual footsteps by dynamically fusing information from multiple frequency components and leveraging human mobility and structural vibration patterns through Bayesian updating; and (3) improving model accuracy by iteratively fusing location information and signal separation. The key novelty in these thrusts lies in fusion of signal processing methods and physical constraints to address real world challenges.
