A Monocular Motion Estimation Algorithm Based on Region Separation

Abstract

The offset of feature points is one of the main reasons for the inaccuracy of the monocular Simultaneous Localization and Mapping Eight-Point Algorithm. In order to overcome the problems of high iteration times and insufficient score in global sampling estimation of fundamental matrix, this paper proposes an improved algorithm for monocular motion estimation. On the basis of ORB-SLAM, a method of clustering and segmenting images is introduced into the Frontend. Firstly, after eliminating mismatches, the Density-Based Spatial Clustering of Applications with Noise algorithm is used to separate the feature points of two images into a number of small regions. Then, according to the size of intra cluster divergence, the small regions are sorted and relabeled. Finally, with our multi regional extraction strategy, the feature points in different regions are selected for Eight-Point Algorithm and the matching degree of the estimated results to all the feature points is detected, with the highest score to be best estimation result. By comparing with the RANSAC method of ORB-SLAM, it is concluded that our proposed algorithm improves the accuracy of motion estimation and reduces the average estimation time of fundamental matrix based on TUM dataset.