Calibration Of A 3D Laser Rangefinder And A Camera And Registration Learning Of Color Point Clouds

In recent years,3D color point clouds,which can simultaneously describe the relative position and color information of objects in a 3D space,have been widely used in digitizing the real world,such as industrial inspection,autonomous navigation,cultural relics protection,and virtual reality.By calibrating the 3D laser rangefinder and the camera,3D point clouds are dyed to obtain 3D color point clouds.When constructing 3D color point clouds in a large scene,the registration method is executed to unify 3D color point clouds in different positions into a same coordinate system.This paper researches the indirect and direct calibration methods of a 3D laser rangefinder and a camera,and fuses 3D point clouds and images into3 D color point clouds.For two adjacent frames of 3D color point clouds,a registration network is deployed for unifying them into a large scene.The main contents of this paper include three aspects:(1)According to the principle of the indirect calibration method,this paper develops an indirect calibration method based on multiple geometric constraints.Firstly,the intrinsic parameters are computed by the camera calibration.Then,the plane-to-plane constraint is presented and combined with the line-to-plane constraint to calibrate the extrinsic parameters.Both the intrinsic and extrinsic parameters are further optimized by a global optimization method.Finally,the 3D laser rangefinder and the camera are calibrated by using the intrinsic and extrinsic parameters.The experiments with synthetic data show that the proposed indirect method is convergent and reliable.The experiments with real data further indicate that this indirect method has good performance to fuse 3D point clouds and images into 3D color point clouds.(2)Since the indirect calibration method is relatively complex and the accumulated error negatively affects the fusion of 3D color point clouds,a direct calibration method based on a fan-shaped calibration board is proposed in this paper.Firstly,geometric feature estimation and global optimization are used to compute the feature points in the laser coordinate system.Then,the corresponding feature points in the image coordinate system are extracted.Finally,the data association between the point cloud and the image is directly established to compute the geometric mapping relationship between the 3D laser rangefinder and the camera.The experiments with synthetic data indicate that the proposed direct calibration method is accurate and numerically stable.The experiments with real data further demonstrate that the proposed method can accurately fuse 3D point clouds and images for both indoor and outdoor scenes with different kinds of geometric structures.(3)In this paper,a supervised learning network is deployed for the registration of 3D color point clouds in a large outdoor scene.Firstly,the KITTI dataset based on outdoor scenes is preprocessed to obtain the 3D color point clouds at different positions.Then,a supervised learning network for registering two adjacent 3D color point clouds is presented.Finally,the network is trained with the preprocessed 3D color point clouds to obtain the rigid body transformation and verify the performance of the proposed supervised learning network.Experimental results show that the proposed method can effectively register two adjacent frames of 3D color point clouds,and has high registration accuracy.