A STATE THRESHOLDING FRAMEWORK FOR ENHANCING DAILY ACTIVITY RECOGNITION IN SMART HOMES

Activity recognition has recently aroused people's wide concern because it provides the house with the ability to represent and recognize residents' behaviors and respond to their needs smartly. However, when using non-obtrusive and pervasive sensors instead of wearable sensors, it is quite difficult to determine the underlying activity due to noisy and uncertain data streams and the difference between users' habits. This paper proposes a State-Thresholding-Based Activity Recognition (STBAR) framework which extends the typical activity recognition framework with automated enhancing algorithm. This algorithm proposed by us identifies activity features in sensor data streams by taking account of two factors: 1) what sensor and its state that can be candidate features, and 2) the state changing threshold measures and filters activity features. Only activity features are enhanced algorithmically the measure and the coefficient of feature enhancing are learned automatically from annotated sensor data, thus leading to an improved representation of activity models. As for recognition accuracy, we evaluate our approach by using a smart home environment database CASAS, and the experimental results reveal that the present framework and algorithm outperform the existing baseline approaches.