Dynamic Real-Time Segmentation and Recognition of Activities Using a Multi-feature Windowing Approach

Segmenting sensor events for activity recognition has many key challenges due to its unsupervised nature, the real-time requirements necessary for on-line event detection, and the possibility of having to recognise overlapping activities. A further challenge is to achieve robustness of classification due to sub-optimal choice of window size. In this paper, we present a novel real-time recognition framework to address these problems. The proposed framework is divided into two phases: off-line modeling and on-line recognition. In the off-line phase a representation called Activity Features (AFs) are built from statistical information about the activities from annotated sensory data and a Naive Bayesian (NB) classifier is modeled accordingly. In the on-line phase, a dynamic multi-feature windowing approach using AFs and the learnt NB classifier is introduced to segment unlabeled sensor data as well as predicting the related activity. How this on-line segmentation occurs, even in the presence of overlapping activities, diverges from many other studies. Experimental results demonstrate that our framework can outperform the state-of-the-art windowing-based approaches for activity recognition involving datasets acquired from multiple residents in smart home test-beds.