MAR-VO: A Match-and-Refine Framework for UAV’s Monocular Visual Odometry in Planetary Environments

In planetary exploration, unmanned aerial vehicles (UAVs) play a crucial role in visual navigation and yet face the challenges of the global navigation satellite system (GNSS) absence and environment degeneration, such as low-texture and nonstructural features. To address this issue, we first propose match-and-refine (MAR), a unified deep learning-based feature matching framework. In the matching stage, it incorporates a retentive self-attention (RSA) module to enhance feature matching accuracy. In the refining stage, local feature windows surrounding the key points are fed into a transformer network to refine the feature point positions and filter out mismatches. Then, we develop MAR-visual odometry (VO), a monocular VO system leveraging MAR to localize in planetary environments using feature extractors such as SuperPoint. To test the effectiveness in such scenarios, we construct a virtual planet-like dataset named Terrain-UAV with 24 sequences including 22 800 images in total, using models built in blender. Experimental results show that MAR achieves competitive results in homography estimation and relative pose estimation (RPE) on benchmark datasets while outperforming Terrain-UAV. In addition, our MAR-VO system yields better absolute pose estimation RMSE than monocular point-based VO systems using traditional features and deep learning methods on Terrain-UAV. Notably, with the MAR framework, our VO system exhibits excellent generality, especially in low-texture planetary scenes.