Research On Monocular Vision Based Metrology For3D Rigid Object

Vision-based metrology is one of the essential problems in computer vision.Supported by the actual application demand for monocular vision based passivedetection and parameter measurement of airborne targets, this dissertation isdedicated to the research of the two major problems in monocular vision basedmetrology:(1) the problem of mathematical inverse estimation of camera'sprojection relationship equation;(2) the problem of establishing the2D-3D featureprojection correspondence from the input image to object's3D model, arethoroughly studied:1) For P3P problem, to solve the problem that the numerical precision andstability of state-of-art closed-form methods are still relative low meanwhileiterative methods can return only one of the multiple feasible solutions, a newsemi-closed method which has high numerical precision and stability is proposed.The new method breaks through from the perspective of lowering the order of thepolynomial equations to be solved, and decreases the order to the ever possiblelowest2by exploiting the "collinear constraint" of all the feasible solutions of P3Pproblem. With the new semi-closed method, all the feasible solutions of P3Pproblem can be retrieved with comparable numerical precision to iterative methods.2) For the pose estimation of3D rigid object when there is no2D-3D featureprojection correspondence from the input image to object's3D model given apriori, aiming at the insufficiency of state-of-art methods in convergence radiusand speed, a new iterative pose estimation method based on general2D-3D contourpoint correspondence is proposed. The new method makes the breakthrough fromthe perspective of explicitly establishing the2D-3D feature projectioncorrespondence between the input image and object's3D model on general contourpoints, and solving the feature correspondence establishing task and the3D poseestimation task iteratively and simultaneously. Evident improvements are gained onconvergence radius and speed performance. 3) Based on the above iterative pose estimation algorithm framework, a newnon-Euclidean multi-feature distance map based3D rigid object pose estimationmethod is proposed. The proposed method concentrates on increasing the accuracyof the2D-3D contour points projection correspondence establishing from the inputimage to object's3D model, and by introducing the concept of non-Euclideanmulti-feature distance map, further improvement on pose estimation convergenceradius is obtained, while without incurring too much additional computationalburden.4) Targeting at the real-time processing requirement of monocular visionmetrology task, a multi-scale space based fast pose estimation method for3D rigidobject is proposed. The new method effectively reduces the computationalcomplexity of the monocular vision based pose estimation system without losingthe required parameter estimation precision, and has been successfully applied tovideo-oriented real-time3D rigid object pose estimation.5) For the object contour extraction problem on which the above proposediterative pose estimation methods rely, a new noise-robust modified Chan-Veseactive contour model and a fast active contour implementation scheme based ondiscretized level set representation are proposed, from the perspective ofsegmentation accuracy and processing speed respectively. The former can handlesevere noise while guarantee the active contour to converge to target edges moreprecisely. The latter improves on segmentation speed performance by a largemargin.