Research On Consistency Information Based Robust Model Fitting And Its Applications

Geometric model fitting,which aims to fit all model instances in given observed data,is a fundamental and critical problem in the fields of computer vision and pattern recognition.It has received a lot of attention in a variety of applications such as 3D reconstruction,motion segmentation,object recognition and tracking,and medical image processing.However,in the complex scenes of practical applications,the observed data usually inevitably contain the outliers caused by noise,measurement errors,wrongly extracted feature points.Due to the existence of outliers,the performance of model fitting algorithms is often seriously degraded,which limits their practicability in industrial engineering problems.In addition,in practical problems,the observed data may also be the multiple structure data that contain multiple model instances,where the data of one model instance act as pseudo-outliers to the other model instances.Consequently,different model instances inevitably interfere with each other during the model fitting process,which further brings challenges to robust model fitting.Therefore,the research on efficient and robust methods for model fitting has an important theoretic significance and application value.To effectively mitigate the impact of outliers and pseudo-outliers on the performance of model fitting methods,this dissertation aims to develop highly robust and efficient methods for model fitting and to effectively improve the fitting accuracy and computational efficiency.To this end,based on the consistency information contained in observed data,this dissertation studies how to capture the reliable relationships between the potential inliers in the data,to improve the performance of model fitting.The dissertation mainly focuses on the four key issues that contain the accurate data representation,the reliable neighborhood relationship construction,the robust model selection,and the high-quality data subset sampling.After that,in the dissertation,four novel methods of model fitting are proposed based on the consistency information contained in observed data.These methods are applied to the tasks of line fitting,circular fitting,homography matrix based plane segmentation,fundamental matrix based motion segmentation,and retinal image registration.The experimental results on several challenging datasets show that the proposed methods can achieve superiority over several state-of-the-art methods.The main research contents and contributions of this dissertation are as follows:(1)To effectively improve the robustness of data representation in the data with a high proportion of outliers,in this dissertation,a robust method called preference consistency based data representation for model fitting(PCMF)is proposed to enhance the robustness of data representation and improve the performance of model fitting.Specifically,a preference consistency based data representation algorithm is proposed to effectively capture the consistency information and preference information,for obtaining more accurate data feature descriptions.According to the differences of information entropy between inliers and outliers during data representation,an entropy threshold based outlier detection algorithm is proposed to effectively remove outliers in observed data.In addition,a clustering based model selection algorithm is also proposed to effectively segment the data(in which outliers are removed)and to robustly estimate the number of model instances and the parameters of model instances.The experimental results show that the pref-erence consistency based data representation algorithm can effectively improve the accuracy and robustness of the model fitting methods.In addition,the proposed model fitting method(PCMF)can achieve promising fitting results in multiple structure data with a high proportion of outliers.(2)In order to effectively deal with the problem of constructing reliable neighborhood relationships in multiple structure data with a high proportion of outliers,in this dissertation,a robust neighborhood consistency based model fitting(NCMF)is proposed to obtain the data with consistent relationships and improve the performance of model fitting.Specifically,firstly,based on the motion consistency information and residual consistency information,a neighborhood consistency preservation algorithm is proposed to generate a reliable neighborhood set for each feature match(i.e.,each data point)of the input data,which is used for data point clustering.Then,based on the consistency information of the data in the neighborhood sets,a neighborhood consistency based clustering algorithm is proposed to effectively distinguish the inliers belonging to different model instances from outliers,and estimate the parameters of the corresponding model instances.In addition,a neighborhood relationship based similarity measurement f-unction is also proposed to effectively calculate the similarities between data points,to further improve the fitting performance.The experimental results show that the proposed model fitting method can achieve promising results in fitting accuracy and computational speed.(3)In order to effectively alleviate the influence of outliers on the accuracy of model selection,in the dissertation,a robust motion consistency guided model fitting method(MCF)is proposed to integrate the motion information contained in the data into the model fitting problem for improving the model fitting performance.Specifically,firstly,an effective motion consistency constraint is introduced to obtain the neighborhood set of each data point(i.e.,each feature match),to alleviate the influence of outliers on the fitting performance.Then,based on the motion consistency of the data in the neighborhood sets,a novel guided sampling algorithm is proposed.The model hypotheses generated by the proposed guided sampling algorithm usually contain the true model instances in data with a high probability,which helps to improve the accuracy of model selection.Finally,combining with the generated model hypotheses and the neighborhood sets,a robust model selection algorithm is proposed to adaptively estimate the number of model instances,and to effectively compute the parameters of model instances from the data containing a large number of outliers.The experimental results show that the proposed model fitting method can achieve satisfactory results in computational speed and fitting accuracy.(4)To effectively alleviate the influence of a high proportion of outliers on the sampling efficiency of data subsets,in the dissertation,a robust triple consistent relationships based model fitting method(TCRE)is proposed to effectively guide the sampling process of data subset and improve the computational efficiency.Specifically,based on a spatial consistency contrast,a series of triples of input data(i.e.,feature matches)are captured.Any two feature matches in each triplet satisfy the spatial consistency relationships.Then,based on the triple consistency relationships,a novel initial data subset selection strategy and a robust data subset update strategy are respectively proposed,to effectively sample the reliable data subsets.Finally,a significant model hypothesis is obtained from the sampled data subsets,and it is used as the estimated model instance.Based on the above two strategies,the proposed model fitting method can effectively obtain the initial data subsets and boost the sampling performance,to improve the computational efficiency of model fitting.The experimental results show that the proposed model fitting method can fast generate significant model hypotheses from the data containing a high proportion of outliers,and it can obtain satisfactory computational speed and fitting accuracy.