A VSLAM for Autonomous Driving Vehicles Based on Point-to-Epipolar Optimization

Most VSLAM uses point-to-point reprojection error to optimize pose, which relies heavily on the number of matching between three-dimensional space points and two-dimensional feature points. Especially when the autonomous vehicle is in intense motion, insufficient 3D-2D matching number will greatly reduce the accuracy of pose estimation and even make pose optimization impossible. Therefore, we proposes a VSLAM for autonomous driving vehicles based on point-to-epipolar optimization. In contrast to the point-to-point reprojection optimization, we optimize the pose by minimizing the distance from the point to the epipolar, which does not require the matching of 3D points with 2D feature points. Experiments on the public KITTI dataset show that our method has better performance than state-of-the-art VSLAM.