Segmentation of accelerometer signals recorded during continuous treadmill walking

This paper describes a method for segmentation of triaxial accelerometer signals recorded during continuous treadmill walking. More specifically, we aim at detecting changes in speed and in incline by analyzing the accelerometer signals recorded on the shin or the waist of the walker. The raw accelerometer signals are transformed either in the time-frequency domain (with a Short-Time Fourier Transform) or in a specific features space (which emphasizes the characteristics of the gait). The transformed signals serve as inputs for change-point detection methods which output a number of estimated change times. Several change-point detection methods are tested, either parametric or non-parametric. In particular, a new change-point detection method is introduced, which takes into account the frequency structure of walking signals. The different signal representations and change-points detection methods are evaluated on a corpus of 24 subjects. An analysis of the obtained results is presented for the two considered sensors (waist and shin).