A Novel Registration Method for Scene Reconstruction Based on Accurate Robot Position and Forward Kinematics

The demand for flexible and adaptable production systems requires efficient automation methods. This flexibility relies on accurate and fast scene reconstruction to create digital representations of various objects or scenes. The contribution of this paper consists of two main parts. First, an automated approach to take depth images from a scene where the sensor's position and orientation are referenced to a robot's base. By mounting the depth sensor on the robot, all images are referenced to the base coordinate system. Second, the proposed registration method utilizes these multi-view depth images and the robot's joint data to combine the different point clouds into a single file. To validate the method, the computation time is compared to state-of-the-art global and local registration algorithms such as random sample consensus (RANSAC) and iterative closest point (ICP). The approach presented in this work achieves similar registration results with less computation time, independent of the number of images, demonstrating the efficiency and robustness. Additionally, the approach enables registration for unstructured surfaces and works independently of overlap between the images. The primary constraint of the proposed method is the requirement for a robotic system with a stationary base and precise knowledge of joint angles. Consequently, this method is best suited for automation tasks involving robotic use while being applicable to both existing and new systems. This new approach significantly enhances both flexibility and accuracy in point cloud registration and enables more adaptable and efficient production systems.