Activity Recognition for Anomalous Situations Detection

Abstract

BackgroundLiving alone can be tough and risky for the elderly, typically, a fall can have serious consequences for them. Consequently, smart homes are becoming more and more popular. Such sensors-enriched environments can be exploited for health-care applications, in particular, Anomaly Detection (AD). Currently, most AD solutions only focus on detecting anomalies in the user daily activities while omitting the ones coming from the environment itself. However, it appears that serious anomalies can be caused by the environment during the user activity such as getting sick during sleeping when it is cold and the window is open. MethodsIn order to consider environmental context with user activities, in this paper, we present a novel approach for detecting anomalous situations occurring in the smart home environment. To that end, we propose as a first step, an activity recognition method based on an hybridization of a knowledge-based technique, taking full advantage of the semantic representation and the reasoning properties of ontologies and a data driven technique based on Dempster Shafer theory. In the second step, given the recognized activity and its surrounding context, we propose an approach that is able to built situations and detect anomalies with a level of uncertainty. ResultsOur system is implemented, tested and evaluated using real data obtained from the Hadaptic platform11http://hadaptic.telecom-sudparis.eu/. and opportunity dataset. The former dataset is used to evaluate the detection of anomalies and the latter is for the recognition of activities. Experimental results prove that with suitable time window size, the activity recognizer and the anomaly detector are efficient having respectively 91% of recognition rate and 100% of precision. ConclusionOur method allows, on one hand, recognizing user activities and, on the other hand, detecting eventual occurrence of anomalies in the user's situation. It proves to be efficient using the tested datasets for each module. However, in order to obtain a more general conclusion we plan to evaluate the method using more different datasets.