Research Of State Estimation Uncertainty In Point-line-based RGB-D SLAM System

Simultaneous Localization and Mapping(SLAM)is one of the key technologies for mobile robots to achieve autonomous motion.Most stateof-the-art feature-based visual SLAM methods use point features to construct data associations.However,under the challenges of fast camera movement,lighting changes,and low-textured scenes,the performance of methods based on only point features decreases,and even tracking fails.Line features are widely distributed in the artificial environment,so it can be used in combination with point features to improve the lack of features in visual SLAM.The core of the SLAM methods is to solve the state estimation problem,and typical state variables include the key frame poses and the positions of the map landmarks.According to the law of backpropagation of covariance,the uncertainty of state estimation can be analyzed.The introduction of line features in the point-based SLAM methods essentially increases the landmarks to be estimated and their corresponding observation information in the original optimization problem.The main work in this paper includes:1.For RGB-D cameras,the paper analyzes and compares the pose estimation uncertainty based two algorithms ICP and Pn P,and explains the difference between the two algorithms in practical applications.2.For the Local Bundle Adjustment,the influence of the number and quality of landmarks on the pose estimation of local keyframes is analyzed,and the corresponding expression of the state estimation uncertainty is derived.3.Based on point features and line features,a RGB-D SLAM system is designed and constructed,in which the line features are used to guide the generation of 3D points and build the corresponding reprojections.The experimental results on two well-known public datasets are consistent with the theoretical analysis of the state estimation uncertainty in this paper,and show that the proposed method has good robustness and localization accuracy.