Reconstruction Based On Corner And Mesh Auto-generation Of Numerical Manifold Method

Reverse engineering(RE) is a new field of CAD/CAM. While conventional engineering transforms engineering concepts and models into real parts, in reverse engineering real parts are transformed into CAD data that computer can recognize or transform real objects into virtual reality environments. There are two key techniques in RE, one is measure, and the other is reconstruction. Normal way of measure is touched-form, including coordinate measuring method and layer analysis method, but it is not proper to measure big object due to poor efficiency and high accuracy requirement of hardware, and layer analysis method makes object destroyed. Recently, with the development of non-touched measure method, computer vision measure method has been popular. Based on computer vision and image processing theory, we can gain 3D information of object through processing 2D image. It is very convenient for big objects to use cameras.In this paper, we match corners between object and image to measure and reconstruct object. We have a deeper research of corner detection, including evaluation, application and classifying. We introduce a new comer detection algorithm —— corner detection algorithm based on clustering matched, from an idea that corners generate between some clustering. We scan an image through using a circular window that be divided into 8 units, then we can find corners after analyze co-relation between each unit and each clustering in image. Input a binary image, compare with MIC and Harris algorithm, this new algorithm only responses false corners in diagonal line, but speed is lower when the circular window is bigger.Camera calibration is a necessary process in computer vision measure. In this paper, we describe the theory of linear calibration, nonlinear calibration and self-calibration. We have assembled a machine vision system pc-based, and developed a software module using TSAI calibration algorithm in four months. We test this module again and again, at last the relative error is under 5 thousandths.