Research On 3D Reconstruction Algorithm Based On Monocular Visual SLAM

With the rapid development of society and the continuous progress of science and technology,people's demand for mobile robots in production and life is increasing,mobile robot technology is gradually moving towards autonomy and intelligence.Simultaneous Localization and Mapping(SLAM)is one of the key technologies for mobile robots.With the progress of technology,the cost of vision sensor is lower and lower and the application is more and more extensive.Vision-based SLAM(v SLAM)has become a focus of researchers.Although there are currently many excellent algorithm frameworks and systems to achieve SLAM problems,building more detailed maps and estimate more accurate camera poses has always been the goal,and v SLAM methods to face poor real-time and high calculations.First of all,this paper focuses on the 3D reconstruction of monocular visual SLAM.The research work includes the improvement of the current excellent SLAM algorithm,the design and implementation of key frame culling algorithm,line segment extraction algorithm and 3D reconstruction algorithm based on the characteristics of straight segment.The research focus is a real-time semi-dense 3D reconstruction algorithm based on ORB-SLAM2 and straight-line segment features.It also needs to design excellent key frame culling algorithms and straight line segment extraction algorithms as the basis of the entire algorithm.This paper designs a excellent keyframe culling algorithm to remove redundant keyframes and an anchor point method to improve the line segment extraction of keyframes.This paper propose a novel reconstruction method based on the monocular ORB-SLAM2 by matching the line segment features extracted from keyframes.Specifically,the novel method builds upon ORB-SLAM2,which first outputs a set of keyframes and their corresponding camera poses and a series of map points in real-time.To reduce redundant keyframes,a keyframe culling algorithm(KRC)is proposed.Then a line segment extraction algorithm is adopted to extract line segments in each keyframe.Finally,by use purely geometric constraints to matching2 D line segments from different keyframes to generates 3D scene model.We choose the TUM RGB-D dataset to test.This dataset contains 39 videosequences is ideal for testing our algorithm.In-depth analysis and thorough evaluation in each stage of the proposed method,the experimental results show that the novel method runs steadily and reliably.It also has strong robustness and can quickly generate an accurate 3D scene online.Finally,a series of tests were carried out on the robot platform we designed,and the test results effectively verified the effectiveness and reliability of the platform.Moreover,the experimental platform was put into actual scenes to test the semi-dense3 D reconstruction algorithm based on monocular vision proposed in this paper.Through evaluation and analysis,the experimental results show that the platform can construct a semi-dense 3D map of the scene,and the entire system performs well and is stable and reliable.