Fast Monocular Visual-Inertial Odometry With Point-Line Features and Vanishing Point Constraints

Visual-inertial odometry (VIO) is widely used in autonomous driving, robots, and drones in global navigation satellite system (GNSS)-denied environments. Current VIO methods primarily rely on point features for tracking in the visual front end. However, point-feature-based methods often perform poorly in challenging scenes, such as illumination changes and low textures. Extracting line features as a supplement from images can alleviate the above issues, but it introduces considerable computational overhead. Furthermore, degeneracy is more likely to occur with line features. To address these, we propose a fast monocular VIO that integrates point-line features and vanishing point constraints. We improve the EDLines algorithm by adaptively adjusting the gradient threshold and designing a short-line merging algorithm, which reduces computational time and yields higher quality line features. In addition, we introduce a parallel line filtering algorithm and estimate vanishing points based on it to solve the problem of degeneracy. Experiment results show that the line feature extraction of our method is much faster than that of visual–inertial odometry using point and line features (PL-VIO), and visual–inertial SLAM with point and line features (PL-VINS), and our method also demonstrates better localization accuracy and robustness in challenging scenes.