In this research, a ‘pose’ comprehends action and sensor position. We assume that the shape of a signal contains pose information, so the poses are classified by the shape of the signal. If the shape of each pose is pre-trained in advance, the pose can be classified to the most similar one among the pre-trained shapes. The similarity checking process is done by the Procrustes shape matching. In this algorithm, the pre-shape is extracted by removing translation and size information, and the two pre-shapes are rotated to the optimal angle where the Procrustes distance of two pre-shape is minimized. After those processes, the shape features are obtained and compared without interruption of other geometric features. In this research, the algorithm is verified by a multi-person experiment of which subjects equip sensors on their chest, hand, thigh, and foot. For the hand case, swing and holding actions are both considered. The pose classification accuracy is over 97% in the experiment. Because of the rotation process in the shape matching, this algorithm is expected to show excellent performance when the attitude of the device changes a lot like in a smartphone.

Keywords: pose classification, Procrustes shape matching, full Procrustes shape distance, PDR, smartphone