Estimation of the User Context on Mobile Computing Devices (Diplomarbeit)


Christian Knapp


Gaining knowledge about what a person is doing and what surrounds him is of great interest for all kinds of applications supplying the user with information. With the knowledge about the users context it is possible to adapt and select the information for the user as well as in which form (e.g. display, audio etc.) it is presented. In this thesis an approach to estimate the user context from the consumer-grade sensor signals of a single smart- phone. The smartphone is allowed to be used and worn in four different daily use positions. The estimation is done by processing and classifying the sensor data with a random for- est classifier. The process of creating and evaluating a user context estimation system is discussed in depth in this work. A general analysis of available sensor information, typical user activity types and common smart phone usage positions conducted to identify the key factors and basic conditions is described. Based on the results of this analysis, several relevant classification algorithms and suitable features are compared and the best performing combination is selected. The implementation of the used data logging application as well as the offline and online user context estimation frameworks which utilize the selected algorithms are explained and evaluated. The evaluation of the implemented components is based on a user study conducted during this work, collecting data from 30 participants (varying in gender, height, weight etc.). The data from this study is used to demonstrate that the proposed method robustly generalizes over different persons and different runs of the same person. A final point of discussion is the transition between different sensor positions that possibly occurs concurrently to different user activities. In contrast to other approaches that aim to classify user activity, this approach does not depend on multiple high-precision sensors worn in different positions on the body. As such, the method is directly applicable to daily-use scenarios that involve a single smart phone carried in typical positions.