OAC: Overlapping Office Activity Classification through IoT-Sensed Structural Vibration

Recognizing human activities in an office has great potential for productivity tracking and health monitoring applications. To do this, various sensing methods have been explored, including vision-based, RF-based, wearables, acoustic-based and sensor fusion. These methods have limitations such as limited range, installation requirements (such as line-of-sight, dense deployment, or wearing on the body), and privacy concerns (e.g. video recording). We present OAC, a room-level Internet of Things overlapping activity recognition system using ambient structural vibration. Our algorithm uses observation-based heuristics in the feature extraction process and classifies multiple activities simultaneously using multi-stage supervised learning. We divide activities into categories based on their overlap potential, allowing us to distinguish simultaneous activities. We evaluate on eight subjects in two locations, showing up to 97% classification accuracy for our first activity category, up to 90% accuracy for our second category of activities, and up to 90% accuracy for our combined overlapping activity combinations.