Research On Simultaneous Localization And Mapping Based On ORB Feature Matching

In recent years,with the rapid improvement of artificial intelligence technology and robot manufacturing technology,mobile robots are gradually applied to all aspects of production and life.Their increasingly intelligent functions and services not only greatly reduce people's labor,but also greatly Enriched people's lives.As a mobile platform that can automatically perform various tasks,intelligent mobile robot is one of the most basic and core functions.When a mobile robot enters a new environment,the primary task is to perceive the surrounding environment,avoid obstacles and make corresponding path planning.To solve these problems,real-time positioning and map construction technology is needed,namely SLAM(Simultaneous Localization and Mapping).).Early robot positioning and navigation mainly relied on laser radar,inertial measurement unit(IMU),distance sensor,etc.With the development of computer vision,the camera is cheap and durable,low power,small size,and can provide a lot of rich texture and Features such as color information are widely used in the SLAM field.At the same time,the depth camera represented by Microsoft Kinect V1 has taken the research of visual SLAM to a new level with its advantages of cost-effectiveness and direct depth.Based on the ORB feature points,combined with some of the advantages and algorithms of the latest SLAM scheme,this paper constructs a set of real-time location and map construction system based on RGB-D sensor.The main theoretical basis and work results of this study are as follows:Firstly,the basic framework of visual SLAM is introduced.The system is divided into two parts:front end and back end,and the working principle and function of each module in the front and back ends are explained.Subsequently,the relationship between camera imaging principle and coordinate transformation is analyzed,and three types of cameras are introduced.Their imaging models are analyzed in detail.Finally,the correction function of camera lens distortion is established,and the Kinect V1 camera parameters used in this study are calibrated.Then,two main methods in the visual SLAM front-end are introduced,which are optical flow tracking method and feature matching method.According to their characteristics,the feature matching method is selected as the method of the system front-end in this study.Subsequently,the SIFT(Scale Invariant Feature Transform),SURF(Speeded Up Robust Features)and ORB(Oriented FAST and Rotated BRIEF)algorithms were explained.Combined with their respective characteristics,the verification was completed by experiments,and ORB was selected as the whole system.Characteristics.In order to eliminate the mismatching pairs in feature matching and obtain more accurate position estimation,this paper uses FLANN(Fast Library Approximate Nearest Neighbors)and Random Sample Consensus(RANSAC)method to utilize four-point method.The obtained homography matrix is used for erroneous matching point culling.In the pose estimation,the 3D-3D ICP point cloud registration method is used.For the ICP solving problem,we construct the least squares problem and solve the pose transformation with the Singular Value Decomposition(SVD)method.Thirdly,the backend of visual SLAM is studied.In the back-end optimization,the key frame concept is introduced,and the pose optimization method combined with graph optimization theory is used.The solution of G~2O graph optimization solver is introduced in detail.process.In order to eliminate the cumulative error of the long-term operation of the system,this paper uses the K-tree dictionary-based word bag model to calculate the image similarity,and judges the loopback through the loopback judgment mechanism.Finally,the SLAM build map model and its use are introduced.At the same time,this thesis carried out detailed experiments on the TUM dataset and gave the experimental results.The experimental results were analyzed from two aspects:real-time and accuracy.