关键词: 视觉SLAM, 双目相机, LightGlue, 误匹配剔除, 深度学习

Abstract: Aiming to address the issue of poor localization accuracy of ORB-SLAM2 in texture-poor and low-light environments,this paper proposes a binocular visual SLAM algorithm integrated with LightGlue.By leveraging self-attention and cross-attention mechanisms,LightGlue significantly enhances the accuracy and speed of feature matching,demonstrating superior performance in challenging conditions such as texture-poor and low-light environments.Additionally,an improved RANSAC sampling strategy is introduced,which employs weighted sampling to mitigate the impact of mismatched point pairs,thereby further enhancing the robustness and efficiency of the algorithm.Experiments conducted on the EuRoC dataset and real-world scenarios show that the proposed algorithm achieves substantial improvements over ORB-SLAM2,with a maximum error reduction of 48.2%and an average error reduction of 26.2%.The localization error is reduced to the centimeter level,indicating higher accuracy in complex scenes.

Key words: visual SLAM, binocular camera, LightGlue, outlier rejection, deep learning