Robust Model Fitting Algorithms Based On Multiple Correlation Information

As an important research topic,robust model fitting plays a fundamental role in various subjects of the computer vision.The aim of robust model fitting is to es-timate all the model instances including desired information in the observed data extracted from images.Robust model fitting is often adopted to the estimation of geometric models,such as lines,circles,homography matrices,fundamental matrices and vanishing points.Additionally,it is also widely applied to the scenarios like lane detection,camera registration and 3D reconstruction.Although many achieve-ments have been made in the last few decades,several challenging issues remain unaddressed.First of all,it is nontrivial to undertake model fitting from the input data that are usually contaminated by heavy noises and outliers.Moreover,the inliers coming from different model instances may interfere with each other,which makes the model fitting even harder.In the case that multiple structures are latent in the data,the data distribution could be very complicated in the sense that dif-ferent model instances in the data may overlap with each other and the inliers from different models are distributed unevenly.The performance of existing model fitting algorithms degrades considerably in these scenarios.To overcome the above difficulties and problems,different kinds of correlation infor-mation are explored in this dissertation.Two robust model fitting methods based on correlation information are proposed for multi-structure data:(1)An correlation based model fitting method via subgraph searching is proposed.The proposed method introduces the correlation graph to represent the correlation between data points and model hypotheses.Based on the correlation graph,the generated model hypotheses constitute the induced subgraphs of corre-lation graph.Then the problem of multiple-structure model fitting based on the principle of maximizing consensus is converted to a problem of seeking a set of in-duced subgraphs within the correlation graph.These induced subgraphs need to be subject to the principle of maximizing the consensus set of vertices.Afterwards,an energy function,which analyzes the information such as data fitting error,spa-tial smoothness and structure similarity,is applied to estimate the number of the model instances.The experimental results show that correlation based model fitting method via subgraph searching can correctly estimate the number and parameters of model instances in data.Moreover,the proposed method is fairly robust to the data with intersecting model instances.(2)A correlation based model fitting method via voting scheme is pro-posed.The core of the proposed method is a hierarchical voting scheme,which thor-oughly analyzes the correlation between data points and model hypotheses.Based on the correlation voting to model hypotheses and data points,the bad hypothe-sis and the outliers are gradually removed.Such that clustering performance on the remaining data points is enhanced in terms of its accuracy and efficiency.In turn,multiple model instances are estimated effectively and efficiently.According to the experimental comparison in various model fitting tasks,the proposed correla-tion based model fitting method via voting scheme can achieve the model fitting in various multiple-structure data and obtain superior fitting results and performance.