Algorithms and design aspects of an automated vision based 3-D surface measurement system

Bibliography: pages 122-132.This thesis reports on the investigation, development and implementation of digital/photogrammetric algorithms into a compatible system for measuring surfaces. Each of the important stages of such a measurement are dealt with in the text. Specifically, these include camera calibration, free network adjustment, location and centering of circular targets, orientation determination, the matching and measuring process and handling of results. The chosen algorithms (existing, modified and/or developed in this work) were all incorporated/designed to form an efficient and usable surface measurement system. Of particular importance was the investigation of determining conjugate (matching) surface points in the multiple images. In this respect a novel multi-image correlation search procedure was designed, implemented and tested. This algorithm provides high accuracy matching methods with suitably close provisional matching positions. A series of tests was carried out to study the performance of the algorithm and the results are presented in this work. Most notable was the method's high reliability when using more than two images, even in image areas with highly repetitive patterns. Multi-image correlation is considerably more robust than "traditional" stereo-correlation procedures. Other system tests performed included: tests on the stability of projected light from two off-the shelf projection devices; a test on the effect of PLL synchronisation of the camera-framegrabber combination of the images; tests on the accuracy performance of different centering techniques and surface measurements themselves. It was found that the off-the-shelf slide projector tested did not provide a stable projection, however an overhead projector which was warmed up for over an hour provided a suitably stable projection. The PLL synchronisation of the camera-framegrabber system produced a noticeable line-jitter (between sequential images) reaching over 0.1 pixels in the most badly affected lines. In a simulated test with artificial targets, template matching obtained the most accurate centre coordinates, however the much faster weighted centre of gravity with grey value as weight technique also provides highly accurate results. These two centering techniques agreed to 1/100th of a pixel when centering with real targets. The much faster centroiding technique is thus highly recommended for any application which requires high processing speeds (such as with on-line systems). Surface measurement precisions of 5/100th mm in the plane of the surface and 15/100th mm in depth were achieved in the measurements of the test objects. These objects all had similar dimensions with a diagonal of about 250mm in length. These accuracies could be substantially improved with higher resolution cameras and more images. Together, the algorithms presented in this work formed a surface measurement software program. The success of many of these algorithms, such as the target location method, and the semi-automatic point identification and exterior orientation determination procedure, could not be gauged with results as such, but by their successful incorporation into the system as properly functioning units