GS-RANSAC : An Error Filtering Algorithm for Homography Estimation based on Geometric Similarities of Feature Points

Augmented Reality (AR) is intended to generate information by displaying augmented objects on real-world objects. AR is essentially used to calculate the coordinates of augmented objects, for which a homography estimation method involving two images is generally used. In homography estimation, the RANSAC (Random Sample Consensus) algorithm is used to select the four most appropriate pairs of feature points extracted from the two images. However, conventional RANSAC algorithms cannot guarantee the geometric similarity of the inter-image locations of the feature points selected randomly. In order to resolve this conundrum, we propose an algorithm to evaluate the geometric similarity of inter-image locations of feature points. The proposed algorithm draws tetragons of feature points on each image. Then the algorithm determines if the tetragons are similar in the order of vertices and the range of internal angles. The experimental results show that the proposed algorithm decreases the failure rate by 8.55% and displays the augmented objects more accurately compared with conventional RANSAC. We improved the accuracy of augmented object coordinates in AR using our proposed algorithm.