[P-PS09] Lunar Science and Exploration

[PPS09-15] A deep learning-based local feature method and its application in photogrammetry on the Moon and Mars

*Jiageng Zhong¹, Jianguo Yan¹, Ming Li¹ (1.State Key Laboratory of Information Engineering in Surveying Mapping and Remote Sensing, Wuhan University)

Keywords:3D reconstruction, image matching, deep learning, rover imagery

Topographic reconstruction of planetary surfaces is crucial for engineering applications and scientific research in planetary exploration missions. Typical reconstruction methods are usually based on photogrammetric techniques, which estimate three-dimensional structures from a set of two-dimensional images. For image alignment, it requires invariant features for correspondence search. Scale-Invariant Feature Transform (SIFT) is one of the most famous and widely used local feature methods. However, hand-crafted features like SIFT perform poorly under weak texture and low light conditions, leading to unsuccessful image matching and the failure of 3D reconstruction. Robust local descriptors based on deep learning outperform handcrafted descriptors, given the greater robustness of convolutional neural networks compared to hand-engineered representations. Therefore, we propose a novel and robust deep learning-based local feature method for planetary terrain reconstruction. We also integrate it with advanced computer vision techniques to establish a 3D terrain surface reconstruction workflow. Through experiments on the images from the Yutu-2 lunar rover and Tianwen-1 Mars rover, we verify the effectiveness and superiority of the proposed method and workflow. It proves to be suitable not only for high-resolution terrain reconstruction but also for other related tasks.