Study On Robust Parameter Estimation Methods Based On Feature Matching

The robust parameter estimation is a critical task and research focus in the fields of computer vision and pattern recognition.In computer vision,the parameter estimations of geometric models are usually based on features,among which point features are the most widely used.Feature matching is the basic work of many computer vision applications,and in essence,it is a problem of mapping between two sets.There are large changes between the images that are about the same scene but obtained in different imaging conditions.Since feature matching is an ill-posed problem,features usually cannot be matched perfectly.Affected by various unfavorable factors,the results of feature matching inevitably involve some outliers.Because outliers do not satisfy the assumed model,they will cause serious interference to the solution of the problem if the parameters of the model are estimated by directly using the matching results with outliers.The traditional methods based on the least squares of errors cannot eliminate the influence of the outliers,and the estimated parameters for the model will seriously deviate from the truth.In order to cope with the outliers in data,the model parameter estimation methods should be robust enough.Therefore,the study of robust parameter estimation methods has important theoretical and practical value.Some key technology of the robust parameter estimation methods based on feature matching is studied in this dissertation.The emphases are put on three aspects: the fundamental matrix estimation,the inlier selection for image feature matching,and the optimization for the robust parameter estimation methods.According to the characteristics of the problems in different situations,several relevant algorithms are proposed,and the effectiveness of the proposed algorithms is verified by theoretical analysis and experiments.Firstly,in order to deal with a large number of outliers in the feature matching results and improve the accuracy of the fundamental matrix estimation,Soft Decision Optimization(SDO)method is proposed.The proposed method takes advantage of the coupling relation between the feature matching and the fundamental matrix estimation,and constructs a soft decision objective function by combining the two processes.The expectation maximization algorithm is adopted to solve the soft decision optimization problem,which can automatically eliminate the influence of the outliers in the point pair set and quickly find the solution to the fundamental matrix.Since the inliers and outliers do not need to be explicitly distinguished during the process of calculating the fundamental matrix,the adverse effects on the fundamental matrix estimation caused by the misidentifications of the inliers and outliers are greatly reduced.SDO combines the feature matching and the fundamental matrix estimation closely in a unified framework,thus significantly improves the accuracy of the estimated fundamental matrix and obtains a higher number of inliers.Secondly,aiming to solve the problems of complex changes between the wide baseline images and the high outlier ratios of the initial matching results,Adjacent Feature Space Consistency(AFSC)algorithm is proposed to select inliers for feature matching of the wide baseline images.AFSC algorithm develops an affine-invariant similarity based on the region area ratio,which is used to measure the degree of topology similarity between two groups of adjacent local feature points.AFSC algorithm selects inliers by two steps,i.e.the adjacent feature space correspondence consistency checking and the adjacent feature space structure consistency checking.AFSC algorithm can quickly select an accurate inlier set from the initial matching results with very low inlier ratios,and its speed is not affected by the changes of the outlier ratios.The proposed algorithm is feasible for the wide baseline images with large differences in viewpoint,scale and rotation.Thirdly,in view of the characteristics of local invariant features,Correspondence Feature Distribution Consistency(CFDC)algorithm is proposed to quickly select inliers for the local invariant feature matching.CFDC algorithm selects inliers by jointly exploiting the distribution consistencies of the features in location,scale and orientation,thus eliminates the impact on feature matching caused by the changes of images in translation,rotation and scale,and effectively improves the precision and recall for the inlier selection.CFDC algorithm adopts the coarse-to-fine strategy,and utilizes Correspondence Feature Classification(CFC)algorithm and k Nearest Neighbor Matching Similarity(k NN-MS)algorithm to perform the coarse selection and fine selection for feature matching respectively.CFC algorithm is a pre-selection algorithm based on One-Class Support Vector Machine(OC-SVM),which can be utilized to select a candidate inlier set from the initial matching results and to achieve the fast coarse matching for the initial matching results.kNN-MS algorithm computes the matching similarities for the corresponding point pairs by the geometric transformation relations of the adjacent feature point pairs,and then performs the fine selection with the candidate inlier set according to the matching similarities.CFDC algorithm has stable performance,fast running speed and high robustness to some large and complex changes between the images.It is applicable to various local invariant features such as SIFT and SURF,which provide the scale and direction parameters.Finally,considering of the problems of existing robust parameter estimation algorithms in speed,accuracy and robustness,Fast Resampling Optimal Sample Consensus(FROSAC)algorithm is proposed by improving RANSAC algorithm.A pre-validation step is added before model validation to improve its efficiency.A spline-based loss function is adopted to evaluate the quality of models more effectively.The inlier set is optimized by iteratively resampling and model validating,thus the number of samples is significantly reduced,and the accuracy and stability of the solution are improved as well.By gradually refining the inlier set according to the dual-threshold,the problem that RANSAC algorithm is sensitive to the setting of threshold is solved.FROSAC algorithm is stable and robust;furthermore,it has obvious comparative advantages in terms of accuracy and running speed,especially when the proportion of outlier in data is higher than 50%.