Abstract: Autonomous terrain relative navigation (TRN) is expected to play a crucial role of upcoming planetary missions. Conventional TRN methods typically rely on landmark detection and image registration, which achieve relative pose estimation of the spacecraft by establishing data association between navigation images and onboard maps. The association problem between measurements and maps is a complex and computationally intensive task, and requires high accuracy in the prior knowledge onboard. Feature matching is a critical task in computer vision, aiming to achieve stable and accurate image registration under real-world varying conditions through robust visual models. This paper addresses the challenge of autonomous optical navigation for future deep space missions and for the first time, RoMa-TRN, an autonomous TRN method based on RoMa feature matching without maps is proposed. For sequential images captured by a monocular camera, the method employs the pre-trained visual model for front-end feature matching and constructs visual odometry. By solving a least-squares problem, it estimates the inter-frame direction of motion of the spacecraft. This measurement can be incorporated into an onboard navigation filter to update the state of spacecraft. Experimental results demonstrate that, compared with other feature matching methods, the proposed approach achieves accurate and robust image registration under varying illumination and viewpoint conditions. Under extreme scenarios with solar elevation angles of 1° and 89°, no significant performance degradation is observed. At an orbital altitude of 10 km, the method achieves a motion direction estimation accuracy of approximately 0.03°, enabling precise direction-of-motion for navigation filter.

Key words: terrain relative navigation, feature matching, visual odometry, deep learning, guidance, navigation and control