Time-Bounded Activity Recognition for Ambient Assisted Living

Abstract

Robust activity recognition in near real-time is a prerequisite for delivering the smartness intrinsic to the pragmatic realisation of smart homes, environments and so forth. Many of the physical devices necessary for equipping a smart home are already available as consumer electronic devices and certified for use by the public. Yet activity recognition remains the preserve of the research community, despite the array of machine learning and other AI techniques currently available. To-date, research has been dominated by the use of pre-segmented data, resulting in the recognition of an arbitrary activity subsequent to its completion. For assistive paradigms dependent on smart technologies, for example Ambient Assisted Living, such approaches are insufficient. The overall objective must be the identification of an activity within an appropriative confidence level as soon as possible after activity commencement. This paper presents a novel approach, Cumulatively Overlapping windowing approach for AmBient Recognition of Activities (COBRA), for near real-time activity recognition, specifically within 10, 30, 60 or 120 seconds of the commencement of an activity. COBRA utilizes an innovative combination of sliding windows augmented with a logistic regression model. The approach is evaluated using the well-established, open, CASAS dataset.