Abstract:  It is difficult to be speedy as well as accurate in feature matching. To overcome the drawback, this paper proposes a feature matching method based on random forest. This method extracts local invariant features in scale space by SUSurE algorithm, then the features and its adjacent pixels are constructed as training samples. In offline, the random forests are trained and a classification model is acquired to deal with the scale, rotation, illumination and perspective changes. In the online stage, the candidate feature points input RF classifier for realtime classification and feature matching. Comparative tests are made between our approach and SIFT. Experimental results show that the method based on RF is generally more robust and faster in the premise of realtime, and is good at accuracy, as well as adjusting to the illumination changes.



 Key words:  feature matching, random forests, SUSurE, local invariant features, SIFT