Inertial Based Terrain Classification

Abstract

Terrain classi cation is an active area of research in robotics and autonomous o -road vehicle driving. However, this problem has barely been studied with the inertial data of human gait. In this thesis, estimation and classi cation of the type of terrain has been performed on which a person has walk. For this purpose we have recorded the gait patterns with inertial sensors of normal human walk. We select six di erent terrain types with variation in hardness and friction. We record accelerations and angular velocities of 40 subjects with two smartphone embedded (MPU-6500) and smart-watch embedded (K6DS3TR) at three di erent body locations (chest, lower back and right ankle). The recorded data was segmented with two di erent approaches namely stripe based segmentation and step based segmentation of signal. Then we computed a total of 200 spatio-spectral features from segmented strides and steps. We trained two machine learning classi ers namely support vector machine and random forest with 10-fold cross validation method for classi cation tasks. We perform classi cation experiments on indooroutdoor, hard-soft, terrain classi cation as well as with class combinations of binary, ternary, quaternary, quinary and senary terrain classes. We achieved indoor-outdoor and hard-soft terrain classi cation results of 97% and 92% respectively. Furthermore, the classi cation results of 96%, 94%, 92%, 90% and 89% were achieved for binary, ternary, quaternary, quinary and senary class classi cation. From these results we can conclude that terrain classi cation can be performed with higher accuracy from 6D accelerations and angular velocities of strides and single walking steps. The results show that stride segmented features perform better than step segmented features. Moreover, the problem at hand can be solved from any of the given sensor location, senor type and invariant types of shoes.