Two Dimensional Point Set Registration Method And Its Applications In Placement Machine

As an essential problem in various practical applications,point set registration is to obtain the point correspondence and suitable spatial transformation between two point sets by using the effective registration methods.In computer vision and pattern recognition areas,point set registration problem is an important and basic problem,and has been widely used in optical character recognition,augmented reality,target recognition,attitude estimation,medical or remote sensing image registration,scene matching guidance,computational chemistry,computational anatomy,etc.As a research hot spot,point set registration has attracted great attention from the researchers and many methods from the perspectives of graph theory,information theory,and feature description have been reported.However,point set registration is essentially a complex combinatorial optimization problem with NPCs,expecially when there are noise,outliers,and non-rigid deformation,the existing methods cannot be directly applied to solve this problem,which greatly limits its practical engineering applications.Based on the above discussions,this dissertation focuses on the matching cost function construction,feature matching,space transformation,and transformation constraints formulation for general point set registration and feature point registration,and puts forward several new point set registration algorithms;As the engineering application study,the visual positioning of components with rectangle pins or circular pins on surface mount equipment based on the point set registration technology is also investigated.The main content is summarized as follows:A unified objective function,that combines distance measurement items,space transformation restriction items and corresponding relationship restriction items,is proposed for point set registration.In order to preserve the topology of point sets,referring to locally linear embedding?LLE?,the k-connected neighborhood is defined and a new structural risk function-local spatial variation constraint is designed.The considered general non-rigid point set registration problem is formulated under the statistical framework and solved with the expectation-maximization algorithm.The proposed registration algorithm preserves the local topology of point sets and combines the global and local variation constraints.In order to further improve the registration performance,the shape context of the point set is introduced into the registration algorithm and serves as the priori knowledge of point correspondences.A two-class classification algorithm based on mixture model of registration error to solve the image feature point registration problem is proposed.In order to distinguish the correct feature matches from the wrong featrue matches,the differences in the distribution of registration errors between the correct and the wrong feature matching point pairs are analyzed.The conditions that the registration error of the correct feature matching point pair belongs to Gaussian,exponential,Laplace,or Rayleigh distribution,and the spatial transformation of the image is rigid transformation,affine transformation,or non-rigid transformation are considered and the corresponding feature point registration algorithms are proposed.In order to improve the performance of these algorithms,especially when the ratio of outliers is high,a local consistency correspondence screening method is proposed based on the local smoothness of images.Considering the advantages of the Gaussian L₂ distance cost function over the Euclidean distance cost function and the log-likelihood cost function,a robust image registration cost function based on Gaussian L₂ distance is designed for the image matching problem.The feature similarity?image feature similarity,point set shape context,etc.?is introduced into the registration algorithm through the construction of corresponding relation matrix.By combining the deterministic annealing algorithm and the proposed two-step update strategy,the designed cost function is solved.Three registration algorithms for rigid transformation,affine transformation,and non-rigid transformation,respectively,are given.The relations between the proposed algorithms and the commonly used registration algorithms are analyzed.Based on the idea of point-set registration,an online coarse-fine positioning algorithm for surface-mount components with rectangular pins is proposed and the corresponding and transformation in point set registration are decoupled according to the implicit conditions in components positioning application.In order to improve the positioning efficiency and positioning accuracy,two kinds of key points are extracted according to the characteristics of rectangular pins and a coarse-fine online positioning strategy is proposed based on the extracted key points.In rough positioning phase,the corresponding relation between these two point sets is obtained by using the similarity of distance feature and shape feature.In fine position phase,the point set registration problem with known correspondence is solved by using least square method.For ball grid array component positioning problem,the solder balls are seen as a constant feature point and the positioning problem is transformed into point set registration problem.In order to achieve the solder balls detection with high efficiency and high accuracy,a coarse-fine two-step positioning strategy,the precise position and rotation angle of the component are obtained.The coarse positioning and fine positioning are achieved through the rigid point set registration method and least squares method,respectively.In order to evaluate the arrangement of solder balls and the quality of solder balls,an overlapping point pair is proposed and the corresponding judgments are given.