Contextual interpretations of user generated acceleration data

Sammanfattning: Modern phones are all equipped with a range of powerful sensors. One such sensor is the tri-axial accelerometer, which measures acceleration in all three spatial dimensions. The accelerometer was initially included to determine the orientation of the phone, i.e. whether it was being held vertically or horizontally (landscape or portrait mode) but it did not take long before game- and gimmick applications started to incorporate the possibilities that come with this technology. Soon following, applications that promote a healthier lifestyle by measuring our physical activity arose.One such application is Affective Health, a personal life log that reflects stress levels with the help of sensors attached to the body of the user. As acceleration data tend to lack intuitiveness, especially when observed without the necessary context, visualizing data is problematic. The task of turning logs of acceleration of the phone into something interpretable and meaningful is explored in this thesis. Resorting to activity recognition was a way of reducing the overwhelming amount of data, which is mostly irrelevant in the eyes of a user. The results of such an algorithm provide layers of information that, together with a condensed version of the raw data, make end results more captivating.In the opening chapters of this thesis, a quick overview of the problem definition and market is given before assessing research in the field of activity recognition by accelerometry, as well as recommendations for creating effective information visualizations. Furthermore, some of the methods applied in later sections are described to demonstrate the scope - what has been covered in literature studies and toseparate it from the novelty of this thesis. Apart from assessing the viability of activity recognition for a non-strict setup with few constraints in sensor position, the main contribution of this thesis is the design of a prototype that features elements from research on how to promote visual memory retention and draws inspiration from what makes contemporary products and services compelling for consumers, in a format that is aesthetically pleasing and well suited for the indented platform. The concept that is chosen as a blueprint for prototyping is driven by the idea that a sense of intelligence and creativity is essential in making accelerometer data meaningful.An iterative process of extending the scope and improving the performance of the classification ends with a validation step, using accelerometer data collected from a total of five contributors performing two sedentary and four dynamic activities, yielding a result of 97.6% correct classifications. Classification is done by a mixture of a KNN (K-nearest neighbor) and a hierarchical model. As the system is put to test using real world data as part of running the prototype, a completely different set of issues that have an impact on the performance of the classifier and subsequently the reliability of the visualization, is identified.