Echtzeitfähige Merkmalsgewinnung von Beschleunigungswerten und Klassifikation von zyklischen Bewegungen

Sensor based activity recognition is a current field of research and in combination with mobile assistance it will open up new possibilities for applications in medical care, elderly and ubiquitous computing. The present thesis develops concepts, approaches and methods for the realization of a system for real-time recognition of everyday like base level activities on embedded devices, focussing on the systematic analysis of the feature space and the selection of appropriate classifiers. In a field test motion data of 26 subjects was collected using a triaxial accelerometer and gyroscopes. 177 features were calculated from the raw data and reduced without noticeable loss of information using a hybrid method. Afterwards four different classifiers were tested on the example data under different aspects. The decision tree showed the best overall performance. Except climbing the stairs all activities could be recognized with an accuracy rate of about 94%. Different window sizes as well as the use of the gyroscopes did not bring a significant improvement of the results, the removal of vectored features, in favour of not being dependent on the attitude, reduced the recognition accuracy about 5-10%. It was shown that real time mobile recognition of base level activities with a single triaxial accelerometer on the hip without individual training is possible even in everyday like scenarios