Remove approach of wrongly matched image based on ORB features

HUANG Li, LI Xiaoming

（School of Mechanical Engineering and Automation, Zhejiang Scitech University, Hangzhou 310018, China）

Abstract: Aiming at the problems of many wrong matches and poor matching accuracy in featurebased image matching, ORB (oriented brief) and RASNAC (RANdom sample consensus) algorithm were researched. On condition that matching was achieved based on hamming distance in ORB algorithm, a method based on improved RANSAC was proposed. After matching based on hamming distance, the preremoving process was added to eliminate a part of wrong matches. Then, further elimination was achieved through RANSAC algorithm. Meanwhile, the computing time was saved by increasing the amount of initial sample and preestimating the sample. At last several images were used to verify the validity. The experimental results indicate that the method can eliminate most wrong matches to improve the matching accuracy, and can suppress the effects of noise. It’s also invariant about rotation. Meanwhile, it guarantees the computation speed.

Key words: feature point matching; ORB feature; wrong matching