An Improved Point-Line Feature-Based Visual-Inertial SLAM Method

To address the challenges faced by existing visual mapping and localization systems in weak or sparse texture environments, such as visual feature degradation, low pose estimation accuracy, and poor robustness, this study proposes an improved visual feature extraction method integrating point and line features. By calculating adaptive dynamic thresholds in real-time to optimize ORB point feature extraction and improving LSD line feature extraction through short line removal, similar line grouping, and merging, the method incorporates IMU (Inertial Measurement Unit) data to achieve tight coupling with point-line residuals in the backend. The system's performance is further enhanced through optimization with a sliding window and marginalization. Experimental results on the EuRoc public dataset demonstrate that, compared to VINS-MONO and PL-VINS algorithms, the proposed approach significantly reduces trajectory error, effectively improving the accuracy and robustness of mapping and localization. These findings underline the method's significant research value and application potential.