On Visual-Inertial SLAM Based on Deep Feature Extraction and Matching

This paper investigates the enhancement of visual-inertial SLAM performance in challenging environments using deep learning technologies. A novel hybrid SLAM system is proposed, integrating deep feature extraction ALIKED and matching techniques LightGlue. Compared to traditional SLAM methods, we achieve high localization accuracy and improve robustness in low-texture and dynamic environments, leading to more stable trajectory estimation and mapping. Extensive experiments in various scenarios demonstrate the effectiveness, versatility, and superiority of the proposed method.