Unsupervised Multi-resident Tracking in Smart Environments
MetadataShow full item record
Aging is a global challenge that our society will face in the next few decades. Smart environment and ambient assisted living (AAL) offer promising technologies to help people stay active, socially connected, and independent into older age. Ambient binary sensors, such as PIR motion sensors, magnetic door sensors, and contact-based item sensors, offer a low cost, easy to deploy and unobtrusive solution to constructing a smart environment. However, the limited ability of coping with multiple residents in a smart environment hinders widespread adoption of the AAL and smart environment technologies. In this dissertation, we investigate the multi-resident tracking problem in smart environments equipped with ambient binary sensors. First, we establish the theoretical foundation of the hypothesis that human daily indoor mobility is predictable. Based on information theory, we derive an upper bound of predictability using sensor events recorded in over 100 smart homes. In addition, we formulate the multi-resident tracking (MRT) problem in the framework of finite set statistics (FISST) and propose a sensor-vectorized MRT solution, sMRT. In sMRT, the resident movement in the smart environment is mapped to a point target maneuvering in a latent measurement space. The sensors are mapped into the vectors of the measurement by mining the spatio-temporal relation exhibited in the recorded sequence. Furthermore, we propose sMRT-ML by introducing an unsupervised learning procedure to derive the model parameters previously treated as hyper-parameters in sMRT. We also introduce a track consistency optimization procedure to reduce target mismatch errors during tracking. Our proposed sMRT and sMRT-ML algorithms are evaluated using sensor events collected in two multi-resident smart homes. The ground truth labels are provided by an external annotator using a visualization tool, ActViz, that we specifically designed for this research. Experimental results shows promising performance of sMRT and sMRT-ML in comparison with baseline algorithms which rely on the smart environment floor plans and sensor locations to associate sensor events with residents.