HTSR-VIO: Real-Time Line-Based Visual-Inertial Odometry With Point-Line Hybrid Tracking and Structural Regularity

Line features in artificial scenes encode geometric information, such as parallelism and orthogonality that can provide excellent observational landmarks for localization and navigation. However, line features extracted in challenging scenes face problems contain duplicate detection and incomplete structural line segment constraints. To address these problems, we propose an efficient visual-inertial odometry with point-line hybrid tracking and structural regularity. First, a novel point-line hybrid tracking algorithm is given, which utilizes matching relations between the current and previous moments of tracked points to support nearest-neighbor line segment tracking. Then, a line-point binding (LPB) residual error that is positively correlated with point feature reprojection error is proposed, and we incorporate this error into the sliding window optimization for pose estimation. Besides, to fully exploiting structural regularity of line features, we extract vanishing points in multiple views to separate vertical lines from spatial structured lines. Spatial consistency constraints between these vertical lines and gravity vector measured by the inertial measurement unit (IMU) are utilized to refine the body pose. Experimental results with the state-of-the-art (SOTA) methods indicate that the proposed method can improve the accuracy of monocular visual-inertial odometry to at least 13.1%.