A Device-Orientation Independent Method for Activity Recognition

This paper describes an orientation-independent method for detecting activities of daily living based on reference coordinate transformation. With the proposed method, a classification model can be trained using data acquired during a specific sensor orientation and applied to other input signals regardless of the orientation of the device. The technique is validated using activity recognition experiments with four different orientations of a single tri-axial accelerometer placed on the waist of 13 subjects performing a sub-class of activities of daily living. A high subject-independent accuracy of 90.42% has been achieved, reflecting a significant improvement of 11.74% and 16.58%, compared with classification without input transformation and classification with orientation-specific models, respectively.