Time Handling for Real-time Progressive Activity Recognition

In a dense sensor-based smart home (SH), a significant challenge is to segment the sensor data stream in real-time to continuously support progressive activity recognition (AR). In this paper, we evaluate an approach that supports the segmentation of the sensor data stream used for knowledge-driven AR. The approach is based on the notion of dynamically varied sliding time windows, where data segments are formally modeled as time windows and ontological reasoning used to infer the ongoing activities of daily living (ADLs). We then present an algorithm that supports perpetual, real-time activity recognition and provide an implementation of both the proposed approach and the ADL ontology it uses. For evaluation, we developed a synthetic data generator and generated a set of synthetic ADLs. In addition, we implemented a real-time activity recognition system and a simple simulator that plays back a synthetic ADL as if the sensors are activated in real-time. We evaluated the real-time activity recognition algorithm, and obtained 99.2% average recognition accuracy on the synthetic ADLs tested. The ability of the algorithm to discriminate sensors that are activated in error was evaluated for selected ADLs and impressive results obtained.