An Efficient EKF-Based Visual Inertial Odometry Using Both Point and Line Features

This paper presents PLS-MSCKF, an efficient stereo Visual-Inertial Odometry based on the Extended Kalman Filter method using point and line features. In some special scenes with obvious lighting changes, repetitive textures, or weak textures, the tracking and matching of point features often fail, and line features with relatively more stable spatial structures can cooperate with the point features to get more information about the scene. We have done several works on using line features in our proposed PLS-MSCKF. Firstly, since degeneracy arises in the process of using line features, we propose a mis-triangulation culling strategy for line features to ensure that the 3D positions of line features used in the optimization process are more accurate and reliable. Secondly, we propose a lightweight loop closure optimization using point and line information to solve the cumulative drift problem in visual-inertial odometry. Experiments demonstrate that PLS-MSCKF has lower localization error than EKF-based SLAM methods on the public dataset EuRoc.