Recognition Models for Distribution and Out-of-Distribution of Human Activities

By monitoring movements and activities, the pro-gression of neurological diseases can be detected. The implemen-tation of such monitoring requires a high level of documentation, which is hardly possible in view of the ever-increasing shortage of nursing staff. In cooperation with two dementia residential communities, we are trying to gradually relieve the burden on nursing staff by developing an approach to automated documentation. In the attempt to recognise activities in the dementia environment, everyday activities can be well recognised using smartwatch sensor technology and machine learning, as shown in previous results from this research group. However, the literature lags behind (can hardly be found in the literature) on how to distinguish an activity from a non-activity, as a person does not perform an activity to be classified at all times. This paper explores a model to solve this problem, taking several approaches: Approach 1: First step classification to distinguish activity < - > non-activity. Second step activity detection using LSTM if activity was detected in step 1. Approach 2: First step differentiation of activity < - > non-activity directly with LSTM. Second step activity detection with LSTM if activity was detected in step 1. Approach 3: Direct distinction of activity < - > non-activity and activity detection with an LSTM. We show the advantages of the respective smartwatch sensor technology, compare the different approaches of our models to the prediction accuracy of the classification of different activities.