Robust Estimation In Computer Vision Based On Kernel Density

Robust Statistical methods were first introduced in computer vision to improve the performance of feature extraction algorithms. One attractive characteristic of traditional robust statistical methods is that they can tolerate up to half of the data points that do not obey the assumed model (i.e., they can be robust to up to 50% contamination). However, they can break down at unexpectedly lower percentages when the outliers are clustered; also, they cannot tolerate more than 50% outliers. We gave the proof of robustness of the MLESAC estimator, and also improved the projection based M-estimator. Then we proposed a new robust estimator based on kernel density. The contribution of this thesis is as follows:● Reviewed kinds of robust estimators both in Statistics and Computer Vision. In Statistics, the estimators are M estimator and least median of squares. In Computer Vision, the Random Sampling Consensus(RANSAC) estimator, Minimize the Probability of Randomness(PAMINPAN), Minimum Unbiased Scale Estimator (MUSE) and Adaptive Least kth Order Squares (ALKS) Estimator, Residual Consensus (RESC) Estimator.● Presented the robustness proof of a robust estimator MLESAC based on Influence Function. The local robustness is described by the bounded Influence Function.● Improved the projection based M-estimator, and apply it to estimate the Fundamental matrix. There are two improvements: make the kernel density adaptive and the inliers scale estimate robust.