Aiming at the problems of insufficient feature extraction and easy tracking loss of feature point-based SLAM systems in texture-less scenarios, as well as improving the initialization accuracy and robustness of the system in challenging scenarios, the line features on the basis of the ORB SLAM3 framework were incorporated and the visual inertia initialization was improved. Firstly, the LSD algorithm and LBD descriptor were incorporated into the front-end visual odometry part for line feature extraction and matching, the point and line feature reprojection error models were established, and the BA method based on nonlinear optimization was used to minimize the reprojection error, while the adaptive factor was introduced to dynamically adjust the weights of the line features. Then the gyroscope bias estimator was constructed by extending the binocular MNEC constraints, adopting the rotation-translation decoupling optimization strategy and introducing the residual evaluation mechanism to ensure the reliability of visual inertia initialization, while the IMU residuals, the feature point reprojection errors, and the line reprojection errors were jointly used as the constraints of nonlinear optimization for the estimation of the camera position. Experiments were conducted in the euroc dataset and real scenes, and the results showed that compared with the pre-improved ORB SLAM3 algorithm, the improved algorithm improved the localization accuracy by 22.9% under the dataset and reduced the offsets by 1.4 m in the real environment, thus verifying the feasibility and effectiveness of the improved algorithm.