摘要 针对机器人SLAM系统，在实际场景或低纹理场景中提取的有效特征点数量少，使得系统初始化效果差和定位精度不高的问题，提出了一种基于点线特征和IMU信息融合的双目惯导SLAM系统（Stereo Visual-Inertial state estimator based on optimized ORB point feature and line feature，OOL-VINS）.首先，对双目视觉进行点线特征的提取与匹配，通过匹配的特征点构建残差模型，并结合松耦合算法实现系统快速且稳定的初始化.然后，利用点线特征以及三角化算法设计了一种更加鲁棒的方法来获取路标点的3D信息，以此来实现系统的位姿跟踪.最后，根据位姿跟踪过程中构建的局部三维地图，并结合滑动窗口的非线性优化对相机位姿进行更新，提高系统的定位精度.实验结果表明，OOLVINS在TUM纹理结构类数据集上能获取更多有效的点线视觉特征，且特征提取耗时为27ms.在EuRoc和TUM-VI数据集上进行初始化实验，实验表明，OOL-VINS初始化更加快速稳定.同样地，我们使用以上数据集进行系统性能的实验验证.结果表明，该系统的平均跟踪帧率为25Hz，在300m的低纹理场景中，定位精度可达0.072m.
Abstract：For the robot SLAM system，the number of effective feature points extracted in actual scenes or low-quality scenes is small，resulting in poor system initialization and low positioning accuracy.In this paper，a binocular inertial navigation system based on the fusion of point-line feature and IMU information（Stereo Visual-Inertial state estimator based on optimized ORB point feature and line feature，OOL-VINS）is proposed.Firstly，the point and line features are extracted and matched for binocular images.The residual model is constructed by matching point and line features，and combined with loosely-coupled algorithm to achieve fast and stable initialization of the system.Then，a more robust method is designed based on the point-line features and the triangulation algorithm to obtain the 3D information of the landmarks，so as to achieve the system′s pose tracking.Finally，according to the local 3D map constructed during the pose tracking process，and combined with the nonlinear optimization of the sliding window to achieve the update of camera pose，which can significantly improve the accuracy of the system.The experimental results show that OOL-VINS can obtain a more effective point-line visual features on the TUM texture datasets，and point-line feature extraction takes only 27ms.The initialization experiment on the EuRoc and TUM-VI datasets show that the OOL-VINS initialization is faster and more stable.Similarly，we use the above dataset for system performance experiments.The results show that the accuracy of the system in low-texture scenes around 300m can reach 0.072m.In addition，the average frame rate of the system during the tracking process is 25 Hz，which contents the real-time requirements.
冯波,刘桂华,曾维林,余东应,张文凯. 双目点线特征与惯导融合的机器人SLAM算法研究[J]. 小型微型计算机系统, 2021, 42(6): 1267-1275.
FENG Bo,LIU Gui-hua,ZENG Wei-lin,YU Dong-ying,ZHANG Wen-kai. Research on SLAM Algorithm of Fusion Robot Based on Stereo Point-line Feature and Inertial Navigation. Journal of Chinese Computer Systems, 2021, 42(6): 1267-1275.
2021, Vol. 42 
