Hazard Detection Algorithm for Planetary Landing Using Semantic Segmentation

Abstract

This paper presents an application of a computer vision technique called Semantic Segmentation to a Hazard Detection (HD) algorithm for planetary landing. Three convolutional neural network (CNN) architectures are trained with binary (safe or unsafe) safety maps and multi-hazard-class safety maps. Their performance is compared together with a replicated state-of-the-art (SOA) HD algorithm from NASA's Autonomous Landing Hazard Avoidance Technology (ALHAT) project. To train and test each method, we prepared a dataset that is rich enough to reproduce the fine hazardous features by developing a realistic Digital Elevation Map (DEM) generator. Our DEM generator produces realistic terrains in arbitrary high resolution and the safeness of each pixel of DEMs is examined by computing the maximum possible slope and roughness for a given lander geometry and landing attitude. The results show that all three CNN architectures perform better than the replicated SOA HD algorithm for noised DEMs. The networks trained with multi-hazard-class safety maps result in better accuracy in terms of mean intersection over union (mIoU) but do not improve pixel accuracy.