High-efficiency Vision And IMU-based Robot Simultaneous Localization And Mapping

Vision-based SLAM(Simultaneous Localization and Mapping)technology aims to help a mobile robot to localize and perceive three-dimentional(3D)map in an unknown environment,the problem that how to ensure the accuracy and efficiency of the localization and mapping in a challenging scene,is the current bottleneck in the SLAM area.Based on the project of the National Natural Science Foundation of China(NSFC),this thesis explores some solutions for chanllenging low-texture,low-illumination,and fast-motion scenes,and proposes a series of vision and Inertial Measurement Unit(IMU)-based SLAM methods or systems that achieve the state-of-the-art(SOTA)performance in terms of accuracy and efficiency.In general,our main contributions are as follows:(1)For the image feature association problem in low-texture scenes,this thesis proposes an improved ORB algorithm that can establish a relatively reliable point feature association.And then by integrating the proposed hybrid reprojection error optimization model of 3D-3D and 3D-2D,this thesis further proposes a novel robust camera pose estimation method.The experiments show the effectiveness of the proposed method,and it can output available pose estimation results while other solutions are prone to fail.(2)For the 3D mapping problem in low-texture scenes,this thesis proposes an improved 3D-NDT point cloud(PD)registration method.First,the improved method solves an initial registration matrix by the contribution(1).Secondly,the method grids the feature PD with a dynamic grid partition algorithm.Finally,the method optimizes the initial registration matrix by minimizing the probability density function of the gridded PD to finish PD registration.The experiments show the proposed method achieves SOTA registration accuracy and can output a relatively subtle PD model.(3)To improve the efficiency of the vision-base d indirect methods,this thesis explores a novel high-efficiency framework design,which,do not extract descriptors between adjacent frames and extract descriptors between non-adjacent frames(keyframes)for image matching.To this end,this thesis proposes a coarse-to-fine descriptor-independent keypoint matching algorithm that can establish reliable keypoint correspondences without descriptors,and further proposes the Fast ORB-SLAM system.The experiments show Fast ORB-SLAM achieve SOTA in terms of accuracy and robustness;compared to famous ORB-SLAM2,it can produce highly competitive localization accuracy while the speed is nearly twice of ORB-SLAM2;Video: https://www.bilibili.com/video/BV1 w T4y1j7hf.(4)Traditional point feature-based SLAM will decrease in accuracy when point features in images are deficient,considering indoor scenes provides rich structural information(line features),this thesis proposes a novel SLAM system based on point and line features.The experiments demonstrate our system produces a SOTA localization accuracy and is able to operate in badly challenging scenes while other solutions are hard to obtain available results;The experiments in real-world show the proposed system is able to recover more subtle 3D PD map model than famous RGBD-SLAM-v2.(5)The above-mentioned vision-based SLAM systems are still hard to work well in extremely challenging scenes,e.g.,for this,based on contribution(3)and(4),this thesis leverages the IMU for the accuracy compensation,and then proposes a high-efficiency VINS(visual-inertial SLAM)system based on point and line features(PL-VINS).The experiments show that PL-VINS can produce high-competitive localization accuracy in challenging scenes;compared to the famous VINS-Mono,it outputs 12%-16% location error less than VINS-Mono with the same pose update frequency;Video: https://www.bilibili.com/video/BV1464y1F7hk;Open-source code:https://github.com/cnqiangfu/PL-VINS.