A Robust Estimator for Structure from Motion Based on Kernel Density Estimation

A robust model fitting technique is presented for recovery of structure and motion from a sequence. The error bound of the inliers is not needed and the outliers are assumed to be randomly (uniformly) distributed. Unlike other methods in the RANSAC family that need some prior information or a scale estimator to select the best consensus set, we estimate the parameters of the model directly based on Bayesian theory and kernel density estimation. Then the estimated parameters can be used directly or used to determine the best consensus set. The advantage of the proposed method is that it can robustly determine the best consensus set without any prior information of the data and need a lower computational cost than other auto-scale algorithms in the RANSAC family. The proposed method is applied to structure from planar motion estimation. The experiments indoor and outdoor have been done to verify the performance of the algorithm and the very promising results are obtained