A Study On Theory And Methods Of Monocular Vision For Large Object

By virtue of its non-contactness nature, low sensitivity to working environment and fastness of data acquisition, vision based measurement, a method of non-contact measurement, has also been widely used in the fields such as machinery manufacture, On-line product detection and remote measurement and control and so on. However currently, it still faces many challenges. This is because the images are usually noisy, low texture, high distortion, which make the current available vision based techniques difficult to select suitable camera models, to calibrate the measurement system, to robustly detect and locate image features, and finally to satisfy the high measurement quality demand. This dissertation focuses on research and development of new vision based methods, the main contributions include:(1) A homography based method, called MECDPSI, is proposed for camera distortion parameter calibration. Our proposed method at first computes the homography from the space plane to the image plane using image feature points around the camera’s principal point, by which image distortion errors and distortion parameters are at first estimated, then a nonlinear optimization step is employed to simultaneously estimate both the homography and the distortion parameters. The effectiveness of our proposed method is verified by both simulation and real image experiments.(2) P3 P problem is an effective method for camera pose estimation from a single image, and its multi-solution phenomenon has been an active research topic in the field. In this work, we show that if the camera’s optical center is close to one of the 3 toroids generated by rotating the circumcircle of the control-point-triangle around each one of its 3 sides, there is always an additional solution with its corresponding optical center lying in a small neighborhood of one of the control points, in addition to the original solution. In other words, there always exist at least two solutions for the P3 P problem in such cases. Since for all such additional solutions, their corresponding optical centers must lie in a small neighborhood of control points, the 3 control points constitute the singular points of the P3 P solutions.(3) A systematical simulation and analysis of the factors affecting the accuracy of visual measurement is carried out. The simulation results show that the appropriate distortion model selection is the most important factor. When the image resolution is high, the effect of image discretization error is negligible. In contrast, image feature detection error could severely affect the accuracy of the estimated distortion parameters, which in turn affect the accuracy of the visual measurement. Such theoretical results could be of reference value in real applications.(4) Hough transformation is a robust line extraction technique and widely used in various applications. It is known that if the Hough space is not properly discretized, multiple lines could vote to the same bin, or the votes of a single line could spread to multiple bins. In this work, the problem of the minimum required bins is investigated under the condition that two line parametersare both equidistantly discredited and each different realizable line in a point grid image should vote to a different bin. The analytical results on the minimum bins are obtained.(5) An automatic detection and matching method of image quadrilaterals is proposed. Firstly, line segments in image are detected using Hough detector, and then preprocessed by merging and deleting the redundant ones. Secondly, initial quadrilaterals are detected by the nearest neighbor search, then verified by deleting erroneous quadrilaterals based on the geometric properties such as harmonic cross ratio. Thirdly, the quadrilaterals are automatically matched by using SIFT descriptors with local and global similarity enforcement. Experimental results on a Crane Bridge end girder demonstrate the feasibility and effectiveness of our proposed method.(6) A prototype vision based system for mechanical equipment measurement is designed and developed, with the development tool VC++ and Matlab 2010 a. The system includesthe following key component units: image preprocessing, feature detection, feature matching, and 3D construction. Using a welding robot arm as the object, the system demonstrates the feasibility and possibility of vision based technology for mechanical equipment measurement.