Activity recognition has become a key component within smart environments that aim at providing assistive solutions for their users. Learning high level activities from low level sensor data depends on several parameters, one of which is the duration of the activities themselves. Nevertheless, directly incorporating continuous duration values into a model is a complex process and may not prove to be very qualitative. In this paper we aim at discretising activity related durations using different clustering algorithms. We explore the possibility of discretising duration data through the use of rudimentary clustering algorithms such as visual inspection to more established methods such as model based clustering. In addition, a probabilistic model is built that predicts both person and activities from the observed values of sensor sequence, time and discrete duration values. Each of the models created is compared in terms of its performance in the prediction of activities. Following analysis of the results attained it has been found that irrespective of the clustering algorithm used for duration discretisation, incorporating the duration information increases the prediction performance. Prediction accuracy was improved by almost 3% when the model was built incorporating durations.
- Activity recognition
- clustering algorithms