Robotic Model Control Of Image Processing

Abstract

A top-down image processing scheme is demonstrated wherein a model of a robot is used to direct and inform 2D image-processing algoriths. A 3D model is constructed with information about the robotic mani pulators, work pieces and camera positions; control signals modify robot model parameters. 2D projections of the 3D model control, where on a video image captured by a frame grabber, important parameters are located. These update the 3D model to inform the operator as to current locations of the robot in the crucial case when commanded positions are in fact not equal to actual positions due to external perturbations! The 2D image processing algorithms are contained within small, localized point-of-interest, POI, regions. The features within these small boxes are rather unitary and simple. Thus the 2D algorithms face very simple tasks in processing these simple features. Such processes as thresholding, automatic setting of thresholds, edge-detection, Sobel difference operators, centroids, ellipsoids and measurements of positions and angles of specific edges and joints of the robot manipulators and work pieces can be carried out rapidly and robustly. This illustrates the importance of top down vision in robotics and in animal vision! In fact, one of the intellectual sources of this model-controlled robotic vision is the scanpath theory of human vision, wherein cognitive models control active looking or scanning by eye movements, that successively place the fovea onto important features of a scene for checking and updating the cognitive model.