CAD-BASED ROBOTIC VISION

Abstract

This thesis presents an approach towards integrating three different technologies, namely computer-aided design (CAD), computer vision and robotics into one unified CAD-based robotic vision system for three dimensional objects. The justification in their integration lies upon their common approach and common use of certain aspects of science and technology and in the drive for greater automation in the automated manufacturing systems of the future. The first component, the computer-aided design system provides a platform to the user for the construction and visualization of three di111ensional objects. It functions as a CAD with the facilities we would normally associate with conventional CAD systems. There is however one i111portant addition to the conventional CAD system. Three dimensional CAD models constructed by our CAD system are also used as the basis for recognizing and locating that particular object. By this, we mean the determination of the position and orientation of the object from grey level images obtained from a camera. To perform the latter, a 3D CAD model of the object is transformed for a set of information which can be used to match the object. We call this information the, vision model. Vision models forms one half of the input basis for the recognition of the object. The other half being the features extracted from the images which is supposed to contain the object. Once a vision model for an object is precomputed, no further computation is required when the recognition system attempts to recognize the similar object in the future and as vision models are precomputed during an off-line phase, fast recognition of the object during the run-time phase is made possible when the information embedded within the vision model is exploited by the vision recognition system. The second component is the vision recognition system. It consists of two main functions. The first involves image understanding and the second involves the matching process. Before any matching can be attempted, some form of intermediate level processing is necessary where image features have to be extracted from images because they form the other half of our matching basis. The strategy we have adopted for our matching has been influenced by Chris Goad's "Fast 3D Model Based Vision" [17]. The general approach used in the recognition process is edge matching where a simple matching algorithm namely depth first search is employed. We know that vision models are used for matching, they are also used by VRS for the selection, classification and ordering of object features to be considered at each stage of the matching process. The outcome is a specialized search tree which exploits the best sequence of matches for object features with image features. The result at the end of a match is the presence or absence of the object and if present, its possible position and orientation relative to the position of the viewer or the camera. The third component is the robotic system. Whilst only a discussion is provided, issues that concern the interface of this component to that of the CAD and vision system is presented. The underlying principle or the approach towards the design or the overall system lies in the concept or off-line programming or precomputation. The idea being if one has the ability to precompute information that will be used during run-time, it is best done, so that during run-time, we are able to exploit the best possible performance for the system.