Real-time line feature tracking using descriptor and epipolar search in monocular visual-inertial SLAM

Abstract

In feature-based visual simultaneous localization and mapping(SLAM), line features have been researched recently due to the advantage of a complementary role for localization even in textureless environments and structural mapping information. However, there were many cases in which tracking of line features was interrupted due to image blur or occlusion, which resulted in the degradation of the SLAM system. This paper proposes a system to continuously match and fuse line features in real-time. Our method selects the candidate lines between separated frames by reducing the range using epipolar constraints and mesh approximation. In this reduced range, it could finally be determined whether the corresponding lines are the same line by calculating descriptor distances and filtering with similarity error. A managing system to take advantage of these extra tracked lines is also implemented. Our method shows the number of tracked lines rises and improves the accuracy of SLAM as a result. The proposed method also helps estimate a more accurate trajectory in a textureless corridor environment than other state-of-the-art algorithms.