Real-Time Visual Inertial SLAM Algorithm Based on Point-Line Feature Fusion

针对室内弱纹理、光照不足环境下基于点特征的视觉同步定位和建图(SLAM)方法存在的轨迹漂移、特征缺失等问题，本文引入具有更丰富的场景结构信息的线特征，提出了一种点线特征结合的视觉惯性SLAM方法。首先，在特征检测中引入优化的双线性插值法，并采用迭代的方法来寻找亚像素角点；并通过金字塔光流和向前向后检测方法提升跟踪的鲁棒性；其次，针对构建雅可比矩阵时残差量不同，采用分步边缘化的方式，提升了后端非线性优化的时间。实验结果表明，改进后的算法在复杂环境中的适应能力和鲁棒性显著增强，相比PL-VINS算法，边缘化速度平均提升30%，平均精度提升18.2%。Aiming at the problems of trajectory drift and missing features in the visual simultaneous localization and map building (SLAM) method based on point features in the indoor weak texture and insufficient light environment, a visual-inertial SLAM method combining point and line features is proposed by introducing line features with richer scene structure information in this paper. Firstly, an optimized bilinear interpolation method is introduced in feature detection and an iterative approach is used to find sub-pixel corner points, and the robustness of tracking is improved by pyramidal optical flow and forward-backward detection methods; secondly, stepwise marginalization is used to improve the back-end nonlinear optimization time with respect to the different amount of residuals when constructing the Jacobi matrix. The experimental results show that the improved algorithm is significantly more adaptable and robust in complex environments, with an average increase of 30% in the marginalization speed and 18.2% in the average accuracy compared to the PL-VINS algorithm.