Population ageing makes providing effective care for the elderly an increasingly significant challenge. Smart home technologies leverage sensors and machine learning methods to monitor residents' activity and support independent living.
In this thesis we discuss the problem of predicting smart home events with the aim of establishing situational awareness. We provide an overview of the field, covering commonly used datasets, activity recognition algorithms, and evaluation metrics. Using the Kyoto dataset from the CASAS database, we evaluate activity recognition using the AL-Smarthome algorithm implemented in Python as part of the CASAS project. Predictions of eight activities are assessed using precision, recall and F1 measure. We plot ROC curves and visualize the predictions on timelines.
Results show that the model detects the majority of true activity segments (macro average recall 0.71), but it also generates a considerable amount of false positives (macro average precision 0.39), which is mostly due to boundary detection issues. We conclude that model performance could be improved by including unlabeled data, using additional machine learning methods for change point detection or using time windows based on the duration of individual activities.
|
