基于地平面的单目视觉辅助激光雷达SLAM研究

Research on Ground-Plane-Based Monocular Aided LiDAR SLAM

融合视觉传感器和激光雷达可以实现优于单一传感器的同时定位与建图(SLAM)系统,现有的视觉和激光雷达融合算法仍然存在计算复杂度高、系统精度及稳定性受错误的深度匹配影响等问题。为了更加高效、鲁棒地融合视觉和激光雷达的信息,充分利用图像和激光点云中的地平面信息,提出了高效的视觉辅助激光雷达SLAM算法。首先,从激光点云中分割出地面点云用于提取图像中的地面ORB特征点,并通过单应性变换中的交比不变性校验特征匹配,从而高效鲁棒地利用单应性矩阵分解实现绝对尺度相机运动估计。然后,将得到的相机运动估计以李群SE(3)形式进行插值,用于校正激光雷达在自身运动过程中产生的点云畸变。最后,单目相机的运动估计作为初值用于激光里程计的位姿优化。公共数据集KITTI和实际环境的测试结果表明,本文算法可以有效利用相机运动估计对激光点云畸变进行校正,实时准确地实现里程计和建图。

The fusion of a vision sensor and LiDAR can achieve a simultaneous localization and mapping(SLAM) system superior to a single sensor. However, the existing vision and LiDAR fusion algorithms still have such problems as high computational complexity and the system accuracy and stability susceptible to wrong depth matching. In order to combine vision and LiDAR information more efficiently and robustly, we made full use of ground plane information in the images and LiDAR point clouds, and proposed an efficient SLAM algorithm of vision-assisted LiDAR. Firstly, the ground point cloud was segmented from the laser point cloud to extract the ground ORB feature points in the images, and feature matching was verified by the cross-ratio invariance in the homography transformation. In this way, the absolute scale motion estimation of camera was realized efficiently and robustly via the homography matrix decomposition. Then, the obtained motion estimate of the camera was interpolated in the form of Lie group SE(3) to correct the point cloud distortion generated by the LiDAR during its own motion. Finally, the motion estimate of the monocular camera was taken as the initial value for the position optimization of LiDAR odometry. The test results of KITTI, a public data set, and the actual environment show that the proposed algorithm can effectively employ the motion estimate of the camera to correct the point cloud distortion of LiDAR and achieve odometry and mapping in real time and accurately.