A Robust and Efficient Visual-Inertial SLAM Using Hybrid Point-Line Features

Visual simultaneous localization and mapping (VSLAM) is a foundational technology in robotics, providing an optimal balance of cost and accuracy. However, existing systems often lack robustness in environments with fast motion, dynamic lighting, or low texture. This letter introduces ML-SLAM, a hybrid visual-inertial SLAM system that combines point-line features with learning-based techniques to improve performance in these challenging conditions. Built on the ORB-SLAM3 framework, ML-SLAM incorporates SuperPoint for adaptive keypoint detection and LightGlue for robust feature matching, along with a novel endpoint-based point-line association strategy to enhance tracking reliability in complex scenes. The system also features hybrid feature-based loop-closure detection and tightly coupled bundle adjustment (BA) incorporating inertial measurements, adapted as standard modules in the ORB-SLAM3 backend to seamlessly integrate the hybrid point-line frontend with the established backend. Extensive evaluations on the EuRoC, TartanAir, UMA-VI, and real-world indoor datasets show that ML-SLAM significantly outperforms state-of-the-art (SOTA) methods, with over 20% improvement in localization accuracy compared to ORB-SLAM3.