<p>Line segments provide additional information about scene characteristics, therefore incorporating them into the scene constraints to improve localization accuracy of the visual-inertial simultaneous localization and mapping (SLAM) system is widely studied. However, the computational effort for point and line extraction and tracking is significant, which limits the real-time performance of SLAM systems. To improve the efficiency and accuracy of the point-line-based SLAM system, we propose a line segment merging strategy to remove redundant line segments and obtain high-quality line features, and an optical flow tracking method for line features combining the Lucas-Kanade (LK) algorithm and Progressive Probabilistic Hough Transform (PPHT) algorithm. In particular, we introduce a Preconditioned Landweber Iteration solver for LK, which utilizes a preconditioner and a corresponding relaxation strategy to improve the conditioning of equations generated by LK, leading to improved speed and accuracy of line feature matching. Additionally, we employ this solver in the solution of the Kanade Lucas Tomasi (KLT) algorithm for point feature matching and sliding window optimization. The experimental results show that our algorithm improves the efficiency of feature matching and localization accuracy.</p>