Leveraging walking inertial pattern for terrain classification

The goal of this work is to illustrate how measurements collected during walking by inertial sensors embedded in the shoes' sole can be used to reveal the underlying terrain type. The final aim is to enable the automatic, real time adaptation of the actuated bottom cushioning of the innovative Wahu shoe for the sake of safety and comfort. For this purpose, the gait patterns of the normal walk of different healthy subjects on four different surface types, with different hardness and friction, are collected offline and represented through the three accelerations' time history. These signals are pre-processed and segmented into two different 'elementary' items, a 'walk' object, made of a sequence of subsequent steps, and a 'mean step' object. In both cases, time and frequency attributes are computed and the most explicative selected through a principal component analysis. A cubic SVM classifier is then trained with the experimental data from multiple walking trials and its performance investigated on different validation sets. Confusion matrices show that the complete 'walk' segment performs much better in terms of prediction power and this is encouraging for further development of the methodology in real time.