Calibration of LiDARs with object detection method

Abstract

Multi - LiDAR calibration, or point cloud registration method is an important subject in the field of 3D target recognition and automatic driving. The objective of this task is to find the transformation matrix from target point cloud to source point cloud, so as to transform point clouds in different coordinate systems to the same coordinate system.

At present, matching methods of point cloud mostly use the whole characteristics of point clouds to find the transformation matrix. However, these methods assume a high degree of geometric similarity between two point clouds. Therefore, these methods are only applicable to short baseline scenarios where point clouds are obtained from LiDARs with short distance and similar pose. For large baseline scenarios where point clouds are taken from LiDARs at a large distance with a large viewpoint difference, these methods can not get good calibration results.

This dissertation propose a new LiDAR calibration method Object4Calib++ for large baseline scenarios using bounding box features. It also applies to point clouds generated by other sensors. This method greatly reduces the dependence on the similarity of point clouds and greatly increases the applicable distance of remote point clouds with incomplete similarity. Therefore, it has great application value in practical scenes. Meanwhile, this dissertation also carries out real world and simulation experiments to verify that the accuracy of our method can meet the actual use requirements. What's more, this dissertation also provides a super-large baseline calibration method CityScaleCalib based on Object4Calib++ and verifies its feasibility in Gazebo simulation environment.