基于特征复用的行星车视觉 SLAM 建图算法

A visual mapping algorithm based on feature reuse in complex environments is proposed to solve the problems of limited mapping range，high mismatching rate and feature resource waste in simultaneous localization and mapping（SLAM）for extraterrestrial exploration of planetary rovers. Image feature outliers are constructed as virtual map points to fully mine and utilize distant and unstructured environmental information. Accurate modeling of distant landmarks is completed through delayed triangulation of historical features with small parallax，and mismatched features are effectively identified and eliminated by epipolar error detection on historical features of candidate map points. The algorithm is implemented based on ORB-SLAM3，and verification experiments are carried out by using the Martian simulation MADMAX dataset and Lunar simulation S3LI dataset. Experimental results show that the mapping quality and localization accuracy of the algorithm are significantly better than those of the SOTA algorithms. The number of map points is effectively reduced，the additional computational overhead is controllable，and the real-time performance of the algorithm is maintained，which can provide reliable technical support for autonomous visual navigation of planetary rovers on extraterrestrial celestial bodies.