A 3D Point Correspondences Uncertainty Aware RGB-D SLAM System

In recent years, much research has been done on RGB-D simultaneous localization and mapping (SLAM) system. Mismatching and depth measurement uncertainty are key factors affecting the accuracy of RGB-D SLAM algorithms. Given that, we propose a 3D point correspondences uncertainty aware SLAM system. Firstly, we conduct ORB feature extraction and matching. Secondly. 3D positions of those pair points are reconstructed by combing depth information with Gaussian Mixture Model (GMM) and outliers are rejected and the initial guess of camera motion can be provided based on LMedS. Under the assumption of Chi-Square distribution, the motion results are further optimized by using Mahalanobis distance and Chi-Square test. Besides, the camera motion trajectory is globally optimized by pose graph. Finally, experiment results prove the proposed method can improve the accuracy of localization and mapping.