Automatic Sensor Data Stream Segmentation for Real-time Activity Prediction in Smart Spaces

Abstract

Recently, human activity recognition and prediction have become important functionalities in ambient-assisted living. Activity inference algorithms detect what task a human undertakes, by analyzing the data stream pattern generated from various Internet of Things (IoT) devices. However, determining how the data stream should be segmented in real-time, referred to as data segmentation, remains as one of the most difficult challenges. In this paper, we propose an automatic data segmentation approach for real-time activity prediction by employing the Jaro-Winkler Distance measurement. Our approach selects a breakpoint of a stream by comparing the Jaro-Winkler distance between the training dataset and the data stream and finding a peak among the variations. The resultant segment also becomes new training data after being tagged; this removes the need to segment the stream data manually for humans. From the experiment based on MIT's smart home dataset collected from a real living environment, our approach shows reasonable performance of 76% accuracy even though the dataset size is relatively diminutive.