Activity recognition for motion-aware pervasive systems

Human activity recognition plays a crucial role in the successful development of pervasive systems, which aim to support people in their daily lives. Hence, it is important to investigate how the knowledge of activities as performed by users can be extracted and used within a pervasive system. The example scenario addressed in this study is the detection of human motion using a single tri-axial accelerometer incorporated in a smartphone. Several classification algorithms were used on preprocessed accelerometer data. While the results showed that the most accurate classification (up to 97.8%) is given by a multiclass classifier based on multilayer perceptron neural networks, this was only achieved at a very high computational cost and with training times of several orders of magnitude longer than the majority of the other algorithms considered. This suggests that a simpler algorithm such as the k-NN, albeit with slightly lower recognition rates, is more suitable for pervasive systems. Some preprocessing aspects were found to cause more of an impact on the recognition accuracy than the actual choice of classification algorithms, in particular the number and type of features extracted from the raw data and the size of the feature extraction windows. Other aspects, inherent to the type of activity being recognised, also affected the accuracy of the classification, resulting on certain activities being more accurately recognised than others. The method adopted was well suited for the recognition of short activities with a defined (and possibly periodic) motion pattern, and not so well for other type of activities, which could be better recognised by incorporating contextual information in the classification system. This thesis was written at the University of Southampton in the United Kingdom under the academic sponsorship of Ed Zaluska as part of the fulfilment of the requirements of the Swiss Joint Master of Science in Computer Science of the Universities of Bern, Fribourg and Neuchâtel. The thesis was supervised by Professor Beat Hirsbrunner and Pascal Brugger (PhD candidate) of the University of Fribourg (Switzerland).