Recognition of motion patterns based on mobile device sensor data

• Main Author: Martin, Juan Jose Sanz
• Other Authors: IVIS, Visualisierung und Interaktive Systeme
• Format: Master Thesis
• Language: English
• Published: Stuttgart, Germany, Universität Stuttgart 2010
• Subjects: Artificial Intelligence Learning (CR I.2.6); Image Processing and Computer Vision Scene Analysis (CR I.4.8); Pattern Recognition Applications (CR I.5.4); Nonnumerical Algorithms and Problems (CR F.2.2); The ''Y''; North Magnetic Pole
• Online Access: http://www2.informatik.uni-stuttgart.de/cgi-bin/NCSTRL/NCSTRL_view.pl?id=MSTR-3010&engl=1

54 pages The new generation of mobile phones contains a series of sensors for measuring acceleration, orientation and positions. Such sensor data have been widely used in the field of wearable computing to recognize postures and activities of persons. However, the recognition of activities based on the sensor data provided by one mobile device is still in its early stage. The scope of this thesis is the development of an Android-based pattern recognition system for the classification of activities like walking, running, climbing stairs and descending stairs. The system uses three dimensional information from sensors for acceleration, orientation and magnetic field. Before the data is processed, the acceleration is normalized using the information provided by the orientation and magnetic field sensors. This normalization transforms the acceleration coordinate system to a new one with the z-axis pointing to the center of the earth and the y-axis to the north magnetic pole. To recognize activities, motion and pose, a decision tree was used as classifier. The tree is trained with input objects (vectors) together with the desired output associated with that input. As input vectors, several features from the data acquired are extracted such as average, variance and range values from each of the three axes (x,y,z) in time domain. There is also another useful feature the energy calculated from the spectrum or frequency domain vector. The results obtained were calculated for windows sizes of 1, 2 and 3 seconds of the normalized acceleration. As a first conclusion, window size of two seconds is the compromise solution since it has the best balanced recognition accuracy. The recognition rates, using two seconds as window size, are over 70% for activities such as running or walking and near 60% for climbing and descending stairs.