For future landing on Small Solar System Bodies (SSSB), lander autonomy and decreasing risk are major system drivers. In this paper, passive vision-based hazard detection is used to achieve a precise, safe and lightweight mission solution compared to heavy and power consuming depth-sensor based solutions (i.e. active detection e.g. LiDAR). The work employs deep learning techniques to identify hazardous and safe landing sites using supervised learning strategies. A high-fidelity synthetic dataset consisting of 43,000 RGB images is created, based on the reference scenario of the Didymos asteroid binary system. An envelope of illumination and boulder environment is generated to increase the robustness of the algorithms. The dataset simulates the images acquired during the soft-landing operation on the asteroid surface. A selection of Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) architectures is trained on the dataset. The networks learn to classify each pixel into boulder or nonboulder classes, according to the accurate labels which are generated together with the RGB images. The deep neural network performances are compared and traded-off according to the accuracy metrics and computational speed. The analysis selects optimal network architecture for the hazard detection task in the asteroid environment. The solution is demonstrated to build a fast and reliable risk map to be fused with additional landing trustworthiness criteria to autonomously determine a safe landing site. In particular, the neural network result is postprocessed according to the navigation uncertainty and the landing footprint to increase safety and explainability of the result. Finally, a qualitative verification is evaluated on real mission images. The result of the work is twofold: (1) provide a representative survey of the semantic segmentation neural networks applied to an asteroid mission dataset and (2) compare an artificial intelligence-based solution for hazard detection with image processing techniques.
DEEP LEARNING-BASED PASSIVE HAZARD DETECTION FOR ASTEROID LANDING IN UNEXPLORED ENVIRONMENT
Proceedings of the 44th Annual American Astronautical Society Guidance, Navigation, and Control Conference, 2022 ; Kapitel : 20 ; 319-334
2024-01-01
16 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch