High accuracy Pose Estimation with Computer Vision

This thesis aims to develop a Computer Vision system that measures with high accuracy the position and orientation of a sample attached to an anthropomorphic robotic manipulator. This system is intended to act as the main sensor in an external control loop, to improve the accuracy in the actual sample positioning system for the neutron diffractometer STREESS-SPEC at the Heinz Maier-Leibnitz center in Garching. The experiments that are carried out in this facility, require an error in position to be less than 50 µm in any direction, while for orientation it cannot be higher than 0.5 degrees in each axis. The actual positioning system consists of a 6-axis industrial manipulator that has a repeatability of 50 µm, but an absolute accuracy of approximately 0.5 mm. It is assumed that the main source of error can be explained by a deficient description of the kinematics of the robot in its control system, which leads to an inaccurate pose estimation of the end effector. Due to this fact, it is necessary to rely on an external measuring system that can perceive deviations both in position and orientation. This information will serve as feedback to compensate the pose of the robot. This thesis is part of the RAPtOr project carried out by FAPS (Institute for Factory Automation and Production Systems) and has the aim of providing a solution for this problem. Two different methods were tested for this research. The first one consists in computing the pose of a square fiducial marker, based on the information provided by a set of two cameras in Stereo Vision configuration. This type of fiducial is proposed by the author and consists in an enhanced version of ArUco markers that allows the use of subpixel algorithms for corner detection. In order to detect corners more accurately, a Super Resolution Deep Neural Network (SRDNN) was built. Different tests determined that the improvement due to this technique was not significant. The second one is also based on Stereo Vision, but in this case, the fiducials are Concentric Contrasting Circular (CCC) markers. These are not internally codified, which makes harder the correlation process between images. Despite this, its small size is advantageous, since it is more practical to attach them to complex samples under test. For the latter approach, it was necessary to develop an algorithm that could scan a sample and generate a 3D map with the relative position between markers. As a consequence that all CCC fiducial are alike and do not have an internal codification, they need to be individually identified by their relative position respect others. A laser tracker with 10 µm accuracy was used to compare the performance of the Computer Vision algorithms. The results of the measurements concluded that both methods are similar in accuracy. The cameras available for testing at FAPS resulted to be insufficient for the task at hand, nonetheless, the results were in accordance with the theoretical analysis. For the target applications a new pair of 20 Mpx cameras, with 16.4 mm diagonal will be used. In combination with 50 mm lenses, these cameras are expected to reach the required accuracy for pose estimation.