Learning models for activity recognition in smart homes

Abstract

Automated recognition of activities in a smart home is useful in independent living of elderly and remote monitoring of patients. Learning methods are applied to recognize activities by utilizing the information obtained from the sensors installed in a smart home. In this paper, we present a comparative study using five learning models applied to activity recognition, highlighting their strengths and weaknesses under different challenging conditions. The challenges include high intra-class, low inter-class variations, unreliable sensor data and imbalance number of activity instances per class. The same sets of features are given as input to the learning approaches. Evaluation is performed using four publicly available smart home datasets. Analysis of the results shows that Support Vector Machine (SVM) and Evidence-Theoretic K-nearest Neighbors (ET-KNN) in comparison to the learning methods Probabilistic Neural Network (PNN), K-Nearest Neighbor (KNN) and Naive Bayes (NB) performed better in correctly recognizing the smart home activities.