Visual–LiDAR SLAM for Rover Navigation in Planetary Environments With Plane Constraints and Global Optimization

Planetary rovers are key remote sensing platforms in space exploration, playing a vital role in scene perception and terrain mapping. The safe operation and data acquisition of rovers depend on precise navigation and localization. However, the absence of a Global Navigation Satellite System (GNSS) on planetary surfaces necessitates autonomous localization. The texture-sparse environments and drastic illumination changes present major challenges for visual-only methods. LiDAR enables direct high-precision terrain mapping independent of lighting conditions. This paper proposes a visual–LiDAR fusion SLAM system for planetary rover navigation and terrain reconstruction. To improve depth estimation in low-texture scenes, LiDAR point clouds are used to recover the depth of visual features. To mitigate the degeneracy of LiDAR odometry in unstructured environments, tightly coupled multi-sensor pose guidance enhances system robustness under extreme conditions. Vertical constraints are strengthened using ground point clouds to suppress drift along the vertical axis. Long-range global pose optimization is achieved through frame-to-map feature matching. Experiments conducted on real-world datasets similar to planetary environments, Erfoud and TAIL, show that our method outperforms state-of-the-art SLAM systems in all four tested sequences. The Absolute Trajectory Errors (ATE), relative to trajectory lengths, are 0.222%, 2.235%, 0.959%, and 2.448%. In terms of mapping accuracy, registration errors between reconstructed and ground-truth point clouds on two trajectories of the Erfoud dataset are 0.854 m and 0.935 m. These results demonstrate the system's effectiveness in achieving long-range, low-drift localization and accurate terrain reconstruction for planetary exploration.