Real-Time Visual-Inertial SLAM with Point-Line Feature using Improved EDLines Algorithm

Many traditional visual-inertial SLAM systems rely on point feature method in front-end odometer module. However, point features may perform bad under blurred images or environments with weak textures. This issue can be solved via introducing line features, due to existing structed features such as line segments usually in artificial structured environments. But many published visual-inertial SLAM systems based on point-line features have poor performance in real-time. This paper proposes a point-line feature based visual-inertial SLAM system by developing an improved EDLines algorithm. We improved the EDLines algorithm by calculating image gradient threshold on different images to extract more efficient line segments in less time. The proposed system is evaluated on public datasets EuRoC, and is compared with the state-of-art SLAM systems including VINS-Mono and PL-VIO. Experiments show that the proposed system can improve the accuracy while ensuring real-time performance.