Data Fusion with a Dense Sensor Network for Anomaly Detection in Smart Homes

Research into assistive technologies for the elderly has been increasingly driven by the rapidly expanding population of older adults in many developed countries. One area of particular interest is technologies that enable aging-in-place, which allows older adults to remain in their own homes and live an independent life. Our work in this space is based on using a network of motion detectors in a smart home to extract patterns of behavior and classify them as either typical or atypical. Knowledge of these patterns can help caregivers and medical professionals in the study of any behavioral changes and enable better planning of care for their patients. Once we define and extract these patterns, we can construct behavioral feature vectors that will be the basis of our behavioral change detection system. These feature vectors can be further refined through traditional machine learning approaches such as K-means to extract any structure and reduce the dimensionality of the data. We can then use these behavioral features to identify significant variations across time, which could indicate atypical behavior. We validated our approach against features generated from human labeled activity annotations, and found that patterns derived from raw motion sensor data can be used as proxies for these higher level annotations. We observed that our machine learning-based feature vectors show a high correlation with the feature vectors derived from the higher level activity annotations and show a high classification accuracy in detecting potentially atypical behavior.