Real-Time Visual-Inertial Odometry Based on Point-Line Feature Fusion

Abstract

To improve the localization accuracy and tracking robustness of monocular feature-based visual SLAM systems in low-texture environments, a visual-inertial odometry method combining line features and point features is proposed, taking advantage of the easy availability of line features in real-world environments and the high accuracy of feature-based methods. The combination of point and line features ensures accurate positioning of the SLAM system in low-texture environments, while the inclusion of IMU data provides prior information and scale information. The pose is optimized by minimizing the reprojection error of point and line features and the IMU error using bundle adjustment. An improved EDlines algorithm is introduced, which incorporates a pixel chain length suppression process to enhance the effectiveness of extracted line features and reduce the rate of line feature misalignment. Experimental results on the public EuRoC dataset and TUM RGB-D dataset show that the proposed method meets the real-time requirements and has higher localization accuracy and robustness compared with the visual SLAM method based on single point feature or the method adding traditional line features.