A fast coarse-to-fine point cloud registration based on optical flow for autonomous vehicles

Point cloud registration is a vital prerequisite for many autonomous vehicle tasks. However, balancing the accuracy and computational complexity is very challenging for existing point cloud registration algorithms. This paper proposes a fast coarse-to-fine point cloud registration approach for autonomous vehicles. Our method uses nearest neighbor sample consensus optical flow pairwise matching resulting from a 2D bird's eye view to initialize the coarse registration. It provides an initial 2D guess matrix for the fine registration and effectively reduces the computational complexity. In two-stage registration, our approach eliminates outliers by utilizing our self-correction module, which improves the robustness without using global positioning system (GPS) information. Point cloud registration experiments show that only our approach can process in real-time (71 ms, on average) while achieving state-of-the-art accuracy on the KITTI Odometry dataset, achieving a mean relative rotation error of 0.125(& LCIRC;) and a mean relative translation error of 0.038 m. In addition, real-road vehicle-to-vehicle point cloud registration experiments verify that the proposed algorithm can effectively align two vehicles' point cloud when the GPS is not synchronized. A demonstration video is available at .