Key feature identification for recognition of activities performed by a smart-home resident

Activity recognition is beneficial for continuous health monitoring of smart-home residents, such as patients and elderly people, living in the privacy of their home. We propose an activity recognition approach apposite for a smart home environment. The observations are obtained through multiple sensors deployed at different locations within a smart home. The activities are represented by the features selected from the received observations. The inconsistent order of performing the activities, infrequent occurrences and the presence of overlapping activities make it challenging to select the features with high class representative ability and inter-class discriminative qualities. We select the key features locally within each activity class, which is least affected by the order of performance and the occurrence of other activities. Next, for association of activities, we solve the existing multi-class problem through a specifically designed binary classification with ranking solution, which learns on the correct and incorrect assignments of activities. A comparison of proposed approach with existing methods in terms of recognition accuracy is presented on publicly available ‘Kasteren’ and ‘CASAS’ datasets, representing a range of overlapping and well separated activities of daily life. Our approach tailored towards a smart home environment demonstrates a better accuracy than existing methods.