An Improved Visual Inertial Odometry in Low-Texture Environment

Recent trends focus on the applications of the autonomous positioning of mobile robots. To solve the limitations of pure vision simultaneous localization and mapping (SLAM) system in unknown environment with few textures, repeated scenes and obstructions, this paper adopts a tightly-coupled visual-inertial odometry (VIO) system as the overall framework. We exploit a method of feature matching by combining both point and line features to improve the system robustly. Compared with point features, line features provide more structured information of the environment and can complete the localization tasks where it is difficult to extract point features. Additionally, the line segments are further filtered and merged by the least squares algorithm for more effective tracking. Based on Plücker coordinates and orthonormal representation, the observation model of line features is given. The experiment results of the Euroc datasets show that our improved system can effectively overcome the low-texture issues and it has higher positioning accuracy. Our method supports both monocular and stereo camera, which makes it possible for more application scenarios.