Is General Purpose Sensing a Pipe Dream?
Tahiya Chowdhury, Murtadha M. N. Aldeer, Justin Yu, Joseph Florentine, Amber Haynes, Jorge Ortiz · 2021
Smart spaces equipped with sensors and combined with learning are becoming increasingly popular for many ubiquitous computing applications. However, application performance depends on many factors, including the number of sensors deployed, the phenomena they measure, their placement throughout the space, and spatio-temporal human-activity distribution characteristics. This work examines the sensitivity of activity-recognition models on the placement and distribution of multi-sensing measurement points in space. We present the results of an empirical study, whereby multi-sensing measurement units are placed throughout a room and train models to identify the type of human activity they observe. Our results show that while our models achieve 95% accuracy, performance degrades with distance and can be highly dependent on unanticipated sensor types. Certain phenomena trigger sensors that are difficult to anticipate a priori, supporting the notion that multi-sensing is useful for general activity recognition and improves robustness. We also show that the pattern of triggered sensors may not be consistent for the same event. Our results suggest designing more adaptive classification models and can guide the design of future cyber-physical systems to capture human activities.