On-line hazard detection algorithm for precision lunar landing using semantic segmentation

Abstract

NASA's Autonomous Landing and Hazard Avoidance Technology project has three core functionalities, namely, Terrain Relative navigation, Hazard Detection and Avoidance and Hazard Relative Navigation. In this paper, we present a real time machine learning based algorithm for Hazard Detection to be deployed during the landing phase. A computer vision technique called Semantic Segmentation is used to classify safe and hazardous landing spots for the spacecraft. Randomly sampled Lunar Digital Elevation Maps (DEM) from the Lunar Reconnaissance Orbiter mission of 2009 are used to train the convolutional neural network (CNN). The ground truth is calculated according to the mission requirements and use existing techniques to calculate slope and roughness. Data augmentation techniques are then used to artificially create additional DEMs by transforming the existing data set. The CNN is validated and testing using similarly sampled DEMs. The results show that CNNs perform well for real time processing of spatially correlated data and hence can be useful for performing Hazard Detection and autonomous landing in future missions.