by YU Yufeng from Pekin University, China
Abstract
As we all know, normal GPS is not suitable for a high-accuracy localization. So camera based algorithms are widely used. For a dynamic environment localization, general visual odometry witch based on feature points cannot give a good result. We humans localize ourselves based on not only eyes, but also knowledge. Visual odometry needs knowledge or assumptions too. One quite simple assumption for on-road environment is that the road is flat and the crosses are right angle bended (see Fig. 1). Using this assumption, the rotation parameters can be settled easily based on line segments without accumulative error. Then the position can be solved using normal odometry method. If we have a sketch map, right angle assumption is not necessary.
by HE Mengwen from Pekin University, China
Abstract
Mobile robots equipped with cheap single-beam LIDARs are well suited for acquiring 3D data of environment. But accurate calibration for data fusion is required. Existing calibration methods for multi-LIDAR require specific calibration target or known environment. We intend to develop a calibration method which requires no specifically placed calibration targets, no special environments, no or few manual operations on software. Our approach uses the multi-type geometry features in outdoor point cloud, such as point, line, plane or quadric. In order to increase its robustness and accuracy, we defined a probabilistic distance measure to combine all types of geometry features into one registration process in same scale.