Methods for engineering symbolic human behaviour models for activity recognition

Context-aware systems are becoming an important part of our everyday life and their ability to accurately recognise the user needs plays a crucial role in their performance. Assistive software would be greatly impaired, were it unable to recognise the current user state, as it would result in inability to correctly assist her. A typical approach in such situations is the employment of probabilistic models that describe the possible states and the probabilities for going from one state to another. Usually these models are handcrafted by the system engineer and the transition probabilities are learned to fit the specific problem. However, in order to build and learn the model, a training dataset has to be collected and annotated which in itself implies finding subjects to conduct an experiment, spending time for repeatedly conducting the experiment, and even more time for annotating it. This makes the building of such models not only expensive but also leads to generalisation problems, as the model is not guided by a domain structure but rather by the underlying sensor readings, which could cause suboptimal solutions. A different approach is to generate the probabilistic model from prior knowledge instead of learning it. One approach to generating probabilistic models could be the usage of human behaviour models that are later mapped onto a probabilistic model and an inference engine is used for estimating the user state. It exploits the additional advantage that the natural way of human thinking is based on causes and effects instead of probabilities. There are corresponding theories that it would be much easier for a system engineer to build a non-probabilistic model. Based on the above assumption, this work investigates the ability of symbolic models to encode context information that is later used for generating probabilistic models. It also analyses the problems arising from such approach and the need of a structured development process for model based activity recognition. As a consequence, the contributions of the work are as follows: (1) it shows that it is possible to successfully use symbolic models for activity recognition in the field of activities of daily living; (2) it provides a modelling toolkit that contains patterns for reducing the model complexity; (3) it proposes a structured development process for building and evaluating computational causal behaviour models. In general, the thesis provides a practical guide to implementing and using symbolic models for activity recognition and proposes a structured process for doing it – something that is often overlooked in the field of activity recognition.