Terrain Classification Using Vibration Sensors: Theory and Methods

To improve the response of an autonomous ground vehicle (AGV) as it seeks to complete an outdoor mission, the control system of the vehicle should vary as it encounters specific terrains such as sand, gravel, mud, or ice. To determine when to switch the control system it is beneficial to have accurate and fast online terrain classification. The earliest terrain classification algorithms were vision-based. However, vision sensors may lead to inaccuracy in certain situations, for example, when the terrain that the vehicle experiences cannot be visually detected due to superficial ground cover such as leaves, dry grass, or even a thin layer of water. Hence, vision-based terrain classification should be accompanied by vibration-based classification, similar to how a human driver perceives the terrain using both sight and feel. This chapter presents both theory and methods for terrain classification using vibration sensors. First, it uses a specific result from terrain classification using 2D structured light sensors to experimentally demonstrate that the signature of a particular terrain is given by the magnitude of the spatial frequency response of the system. Next, it is shown that the speed of the vehicle and the vibration transfer function of the system define a map from the spatial frequency response to the frequency responses of the vibration sensors. Hence, the magnitudes of the latter frequency responses serve as speed-dependent terrain signatures. An algorithm based on a Probabilistic Neural Network is then developed for terrain classification using this vibration signature. The online computational burden of this algorithm is substantially reduced by using Principal Component Snalysis to obtain a lower dimensional signature vector. The resulting algorithm is demonstrated using data