Automatic multi-camera hand-eye calibration for robotic workcells

Human-robot collaboration (HRC) is an increasingly successful research field, widely investigated for several industrial tasks. Collaborative robots can physically interact with humans in a shared environment and simultaneously guarantee an high human safety during all the working process. This can be achieved through a vision system equipped by a single or a multi camera system which can provide to the manipulator essential information about the surrounding workspace and human behavior, ensuring the collision avoidance with objects and human operators. However, in order to guarantee human safety and an excellent working system where the robot arm is aware about the surrounding environment and it can monitor operator motions, a reliable Hand-Eye calibration is needed. An additional improvement for a really safe human-robot collaboration scenario can be provided by a multi-camera hand-eye calibration. This process guarantees an improved human safety and give the robot a greater ability for collision avoidance, thanks to the presence of more sensors which ensures a constant and more reliable vision of the robot arm and its whole workspace. This thesis is mainly focused on the development of an automatic multi-camera calibration method for robotic workcells, which guarantees ah high human safety and ensure a really accurate working system. In particular, the proposed method has two main properties. It is automatic, since it exploits the robot arm with a planar target attached on its end-effector to accomplish the image acquisition phase necessary for the calibration, which is generally realized with manual procedures. This approach allows to remove as much as possible the inaccurate human intervention and to speed up the whole calibration process. The second main feature is that our approach enables the calibration of a multi-camera system suitable for robotic workcells that are larger than those commonly considered in the literature. Our multi-camera hand-eye calibration method was tested through several experiments with the Franka Emika Panda robot arm and with different sensors: Microsoft Kinect V2, Intel RealSense depth camera D455 and Intel RealSense LiDAR camera L515, in order to prove its flexibility and to test which are the hardware devices which allow to achieve the highest calibration accuracy. However, really accurate results are generally achieved through our method even in large robotic workcell where cameras are placed at a distance d=3 m from the robot arm, achieving a reprojection error even lower than 1 pixel with respect to other state-of-art methods which can not even guarantee a proper calibration at these distances. Moreover our method is compared against other single- and multi-camera calibration techniques and it was proved that the proposed calibration process achieves highest accuracy with respect to other methods found in literature, which are mainly focused on the calibration between a single camera and the robot arm.