Line Segment Tracking Assisted by Optical Flow for Visual-Inertial SLAM

Point features have poor robustness in scenarios such as sparse texture and image repetition. Therefore, line feature methods of visual simultaneous localization and mapping (SLAM) are leveraged to improve the performance of point features in feature tracking process. However, line features may not be tracked when encounter false matching and misidentification. In this paper, we propose a line segment predicting and neighborhood matching method assisted by optical flow for visual-inertial SLAM algorithm to attempt to solve tracking problem. Firstly, optical flow algorithm is introduced to assist line segments matched by LBD algorithm when the match lines are insufficient. Then, we optimize the body pose by minimizing the marginalization error, visual residual and IMU preintegration residual. Afterwards, the cumulative error is further reduced by loop constraint. Finally, our proposed method is performed on public datasets and achieves high accuracy.