Lunar polar exploration missions impose severe challenges on autonomous visual navigation, including extreme lighting with persistent shadows, high-contrast terrain, and stringent onboard computational constraints. To address the inherent trade-off between navigation precision and real-time performance, this paper proposes a novel two-stage (offline-online) visual navigation framework. In the offline phase, a landmark database is pre-constructed by analyzing shadow distributions under polar illumination conditions, ensuring feature robustness. In the online phase, a rasterization-reflectance rendering method synthesizes high-fidelity terrain images from the database in real-time, which are then matched with actual descent imagery using a masked Normalized Cross-Correlation (NCC) algorithm for precise pose estimation. The framework is validated through both simulations using orbital data and a helicopter-based flight test in a lunar-analog environment. Results demonstrate that the method achieves horizontal position errors within 20 meters and altitude errors within 10 meters, while maintaining an average processing latency of 500 ms per frame on an embedded system. This work provides a practical solution balancing accuracy, robustness, and computational efficiency for future lunar polar landings.
Wang et al. (Sun,) studied this question.