| This thesis examines the complex problem of robotic interaction with moving objects exhibiting few distinctive visual features in the context of marking surface deformation defects for quality control in the automotive industry. The designed pose and motion estimator, which is the central component of the proposed robotic tracking and marking station, embeds a feature-based tracking approach, which builds upon the selection of a limited, but consistent set of features and their tracking on a frame-by-frame basis.;The knowledge provided to the system by the macro-features is successfully integrated into the inter-calibration procedure between the defects detection stage, whose description remains beyond the scope of this thesis, and the autonomous robotic tracking and marking station. As a result, only a limited number of macro-features are sufficient to inter-calibrate two sensing devices, located in two different stations along an assembly line. Additionally, this inter-calibration procedure is performed on-line and does not require a target object.;Also, with the integration of the supervisory layer, the experimental validation demonstrates the robustness of the proposed pose and motion estimator to a series of realistic situations, such as occlusions from the robot, slight changes in the illumination or the reflectance properties of the panels' surfaces, as well as the sporadic appearance of factory associates in the view of the acquisition system.;Different defects marking procedures are tested with an actual robot arm, including a stamping operation on a static object. An experimental validation of the robotic marking operation on a moving panel, using an LED-pointer to mimick a spray-gun end-effector, is also performed. The accuracy achieved in the two marking validation phases demonstrates the suitability of the proposed robotic solution to become a viable alternative to perform fully automated region marking of deformations over large surfaces and for substantial volumes of production.;While the visual acquisition system relies on low resolution cameras, the proposed algorithm provides sub-pixel accuracy on the pose estimation of an automotive panel, and its associated motion. The pose and motion estimator embeds classical computer vision algorithms for feature extraction, matching and tracking. Their limitations, in the case of tracking industrial objects with few contrasting features, are solved from a software perspective, without complicating their mathematical foundations or the hardware architecture of the visual acquisition system. In order to reliably solve the limitations imposed by the general appearance of the objects, coupled with the complex factory environments in which they exhibit their motion, the pose and motion estimator incorporates a supervisory layer, whose goal is to provide time-efficient, accurate and fault-tolerant visual servoing data to the robotic station. The only knowledge provided to the supervisory layer is related to a limited number of macro-features, which are pre-selected over the structure of the automotive panels, when configuring the robotic tracking and marking station. |
