Asteroid Pose Estimation Method Based on Monocular Vision and Intelligent Perception

ABSTRACT

To address the challenges of relative pose estimation in asteroid exploration missions, such as sparse surface texture, dramatic illumination changes and dynamic occlusion, this paper proposes an end‐to‐end pose estimation method based on a geometrically aware dual‐branch network. First, a geometric perception dual‐branch decoupled regression network is designed, utilizing ResNet‐18 as the backbone. Linear fully connected layers are employed to independently regress the rotation and translation parameters, effectively mitigating the parameter coupling problem in multi‐task learning while preserving the physical scale consistency of the translation parameters. Second, a physics‐constrained joint loss function is constructed by explicitly incorporating geometric consistency and angular periodicity constraints, thereby enhancing the prediction accuracy of the model. Simulation experiments conducted on the Bennu dataset demonstrate that the proposed method achieves high‐precision relative pose estimation for asteroids, with a position estimation error of less than 2.80% and an attitude angle accuracy of 1.764°.