A Robust Deep Learning Model for Terrain Slope Estimation

Abstract

Interest in autonomous robots has grown significantly in recent years, motivated by the many advances in computational power and artificial intelligence. Space probes landing on extra-terrestrial celestial bodies, as well as vertical take-off and landing on unknown terrains, are two examples of high levels of autonomy being pursued. These robots must be endowed with the capability to evaluate the suitability of a given portion of terrain to perform the final touchdown. In these scenarios, the slope of the terrain where a lander is about to touch the ground is crucial for a safe landing. The capability to measure the slope of the terrain underneath the vehicle is essential to perform missions where landing on unknown terrain is desired. This work attempts to develop algorithms to assess the slope of the terrain below a vehicle using monocular images in the visible spectrum. A lander takes these images with a camera pointing in the landing direction at the final descent before the touchdown. The algorithms are based on convolutional neural networks, which classify the perceived slope into discrete bins. To this end, three convolutional neural networks were trained using images taken from multiple types of surfaces, extracting features that indicate the existing inclination in the photographed surface. The metrics of the experiments show that it is feasible to identify the inclination of surfaces, along with their respective orientations. Our overall aim is that if a hazardous slope is detected, the vehicle can abort the landing and search for another, more appropriate site.