In Active and Assisted Living (AAL) systems, a major task is to support old people who suffer from diseases such as Dementia or Alzheimer. To provide required support, it is essential to know their Activities of Daily Living (ADL) and support them accordingly. Thus, the accurate recognition of human activities is the foremost task of such an AAL system, especially when non-video/audio sensors are used. It is common that one or more sensors could share or represent a unique activity and the estimation of the most optimal window size for such activity is challenging. Motivated by the powerful learning ability of neural models architectures, this paper proposes to bridge dynamic windowing and Recurrent Neural Networks (RNN), which results in producing the estimated window of sensor events and recognizing the related activity, consequently. The proposed RNN model is trained based on a dynamical systems perspective on weight initialization process. In order to check the overall performance, this approach was tested using the popular CASAS dataset and the newly collected HBMS dataset. Compared to other approaches, the results show a high performance, based on different evaluation metrics. We believe that the proposed windowing approach and RNN model can assist to detect subject independent human activities in smart environment.

Top