Activity recognition based on inertial sensors for ambient assisted living

Ambient Assisted Living (AAL) aims to create innovative technical solutions and services to support independent living among older adults, improve their quality of life and reduce the costs associated with health and social care. AAL systems provide health monitoring through sensor based technologies to preserve health and functional ability and facilitate social support for the ageing population. Human activity recognition (HAR) is an enabler for the development of robust AAL solutions, especially in safety critical environments. Therefore, HAR models applied within this domain (e.g. for fall detection or for providing contextual information to caregivers) need to be accurate to assist in developing reliable support systems. In this paper, we evaluate three machine learning algorithms, namely Support Vector Machine (SVM), a hybrid of Hidden Markov Models (HMM) and SVM (SVM-HMM) and Artificial Neural Networks (ANNs) applied on a dataset collected between the elderly and their caregiver counterparts. Detected activities will later serve as inputs to a bidirectional activity awareness system for increasing social connectedness. Results show high classification performances for all three algorithms. Specifically, the SVM-HMM hybrid demonstrates the best classification performance. In addition to this, we make our dataset publicly available for use by the machine learning community.