1 Department of Informatics and Mathematical Modeling, Technical University of Denmark
This dissertation is concerned with the introduction of vision-based application s in the ship building industry. The industrial research project is divided into a natural seq uence of developments, from basic theoretical projective image generation via CAD and subpixel analysis to a description of an implementation in real production environments. The theory for projection of world points into images is concentrated upon the direct linear transformation (DLT), also called the Extended Pinhole model, and the stability of this method. A complete list of formulas for calculating all parameters in the model is present ed, and the variability of the parameters is examined and described. The concept of using CAD together with vision information is based on the fact that all items processed at OSS have an associated complete 3D CAD model that is accessible at all production states. This concept gives numerous possibilities for using vision in applications which otherwise would be very difficult to automate. The requirement for low tolerances in production is, despite the huge dimensions of the items involved, extreme. This fact makes great demands on the ability to do robust sub pixel estimation. A new method based on cross correlation is presented. Working with vision in harsh environments with few possibilities for controlling light, vibrations, electrical noise etc. requires knowledge about all factors and components in the vision sys-tem, which can possible influence the image generated. A description of the experienc e achieved during the project is provided. The project is industrial oriented. An essential part of the project has been focused on the possi-bilities for immediate use of the results. A full implemented application doing vision ba sed positioning is described. It is concluded that visionbased applications in ship building are not only pos sible, but also holds great potential in the area of quality control and automation. The effort involved is not necessarily very great, at least not, if the vision-based information can be accompanied by relevant CAD information.