Matthias Putz Fraunhofer IWU Chemnitz, Germany Mohamad Bdiwi Robotics Department, Fraunhofer IWU Open-Box Target for Extrinsic Calibration of LiDAR, Camera and Industrial Robot Aquib Rashd Robotics Department, Fraunhofer IWU Chemnitz, Germany Aquib.rashid@iwu.fraunhofer.de Wolfram Hardt Department of Computer Engineering, TUC Chemnitz, Germany Alexey Kolker Faculty of Automation and Computer Engineering Novosibirsk State Technical University Russia Abstract—Low cost 3D LiDAR complement cameras in perception application for various industrial environments. Safe and efficient human robot collaboration requires easy and accurate extrinsic calibration of sensors with an industrial robot. This work presents an efficient and accurate method for extrinsic calibration between LiDAR, camera and a heavy-duty industrial robot. Open-box target mounted on robot enables parameter estimation by constraining sensor data to multiple planes, which constitute the target surface. The method enables eight correspondences between the sensors and the robot in each data sample. The method enables speedup in sensor setup and drastically reduces efforts required for data collection through automation. The results have been evaluated for simulated and real environment. Keywords-Extrinsic-calibration; LiDAR; Camera; Industrial Robot, open-box, plane constraint I. INTRODUCTION Applications of low-cost 3D LiDAR in production industry for human robot collaboration is an immerging field. 360° horizontal field of view can help detect human entry as well as track worker inside the cell to increase flexibility and safety [1]. Fusion of 3D LiDAR and camera is being used for various robotic perception research to improve flexibility and efficiency [2, 3]. Multi-modal sensor fusion enables higher estimation of dynamic object detection. This enables collision avoidance with large heavy-duty robot (manipulator arm). Extrinsic calibration; which is required for fusion; deals with estimating pose of one sensor with respect to other. Sparse LiDAR data of less than 16 channel makes the accuracy of these estimations challenging. Inaccurate estimations may lead to wrong shortest distance estimation between robot and human, which in turn may result to collision. Furthermore, in an agile production cell, sensor position may change often based on the process specifications. Thus, faster and accurate extrinsic calibration is of high importance. The accuracy of calibration estimation depends on the quality and quantity of point correspondences. These properties are directly related to the sensor used, e.g. field of view and resolution of sensors. Furthermore, the method used to estimate the 3D-3D or 3D-2D point correspondences between sparse LiDAR and high-quality camera data. Although there exist various methods for LiDAR and camera calibration, the field of joint calibration of low-resolution 3D LiDAR, with camera and an industrial robot is an active field of research. Available methods can be categorized into target-less [4- 7] and target-based approaches [8-13]. The target-based approach can provide required accuracy and efficiency for stationary sensor setup. The available methods [8-13], however, are inefficient for the extrinsic calibration of low- resolution LiDAR with camera and industrial robot, with constraints on data collection within robot cell. Calibration methods for low resolution LiDAR focus on filtering and fitting the sparse data for accurate estimations. Park et al. [10] emphasized on noise generated in LiDAR data from traditional checkerboard targets, and proposed using white triangle and diamond shaped planer targets with unbiased inlier ratio to improve plane constraints. Furthermore, by using virtual points in their 32-channel LiDAR data, the method enhanced the polygonal edge estimation, which in turn resulted in higher accuracy corner estimation. These corner positions acted as point correspondences between LiDAR and manually clicked image corner pixels. Dhall et al. [13], used similar diamond shaped target boards, as by [10], for extracting 3D point correspondences in a 16-channel LiDAR. This method, furthermore, used ArUco markers [14] to remove manual image correspondences selection. Pusztai et al. [11] proposed flexible method for extrinsic calibration by using ordinary boxes. The only constraint while collecting data is to have three perpendicular sides visible. The method proposed filters a 16-channel LiDAR data using box fitting and refinement methods. This results in extracting seven-point correspondences from each data sample.