Robust Multi-scale ORB Algorithm in Real-Time Monocular Visual Odometry

Abstract

In this paper, a novel multi-scale ORB algorithm with lower computation is proposed applied for increasing correct matches of feature points in visual odometry when image scale changes. Since ORB algorithm has little scale invariance for feature points matching, the visual odometry employing the ORB algorithm directly performs poorly in the position and orientation estimation. Therefore, the proposed algorithm combines the ORB with SURF by added the scale space. In addition, single layer non-maximum suppression is applied to the selection of stable feature points to decrease spending time in matching step. Experimental results present that the proposed algorithm achieves good matching performance in terms with scale invariance taking into consideration. It was found that the position estimation and the orientation estimation was improved compared to the visual odometry based on the ORB algorithm while the spend of time has only increased a little.