Fusing multi-modality sensors enriches the dimensions of the robots' perception information, which can effectively improve the performance of several perception tasks, such as object detection, depth complement, and mapping, etc. To jointly use these raw data from different sources, the coordinate systems of different sensors need to be aligned spatially and temporally. In this thesis, we focus on the issue of extrinsic calibration for 3D LiDARs-cameras system and propose three different methods, which contain target-based and targetless method. In the target-based method, two novel algorithms are proposed for sparse LiDARs-cameras system. The extrinsics in the first method are calibrated by the artificial corner target, which consists of three orthogonal surfaces with fiducial markers...[ Read more ]

Fusing multi-modality sensors enriches the dimensions of the robots' perception information, which can effectively improve the performance of several perception tasks, such as object detection, depth complement, and mapping, etc. To jointly use these raw data from different sources, the coordinate systems of different sensors need to be aligned spatially and temporally. In this thesis, we focus on the issue of extrinsic calibration for 3D LiDARs-cameras system and propose three different methods, which contain target-based and targetless method. In the target-based method, two novel algorithms are proposed for sparse LiDARs-cameras system. The extrinsics in the first method are calibrated by the artificial corner target, which consists of three orthogonal surfaces with fiducial markers and retro-refective tapes. This calibration can be finished in one shot, which is suitable for massive production. The second one takes the checkerboard as the target, which is flexible and easy to implement. Both methods utilize plane-plane correspondences to initialize the extrinsic parameters and further optimize them by edge alignment. In the targetless method, the proposed method utilizes the feature points derived from a prominent plane in the scene and iteratively minimizes the reprojection error. A maximum likelihood estimator (MLE) is designed by considering the uncertainty information of the measurements. Furthermore, we explore the distribution of collected data and characterize the robustness and solvability of the extrinsic estimates using a confidence factor. All the procedures in our approaches can be automatically operated without manual intervention. Simulation and real-world experiments both qualitatively and quantitatively demonstrate the robustness and accuracy of our methods. The comparison experiments show that the proposed methods outperform other methods.