WEMA-VINS: Visual-Inertial SLAM With ELSED-MIS Line Filtering

Abstract

Unmanned aerial vehicles (UAVs) are widely used in various fields due to their flexibility and maneuverability. However, for UAV systems that rely solely on point features, weak-texture environments cause a decline in pose estimation accuracy, creating a severe localization challenge. The fusion of line and point features introduces more geometric constraints, enhancing system robustness and localization accuracy. This letter presents a lightweight monocular visual-inertial SLAM framework, WEMA-VINS (Weak-grid ELSED MIS-filtered and Angular-constrained VINS), designed for robust point-line fusion. It integrates weak-texture region elimination, ELSED line detection, Maximum Independent Set (MIS)-based pruning of redundant segments, and an angular-constrained line optimization module. Our proposed framework collectively improves point feature quality, preserves structural line information, and enhances robustness against endpoint drift and partial occlusion, leading to reliable point-line fusion and efficient state estimation. Extensive experiments on the KAIST VIO and EuRoC datasets demonstrate that WEMA-VINS significantly improves both localization accuracy and robustness, consistently achieving leading performance over recent point-line SLAM baselines while maintaining real-time efficiency.