Research On Mobile Robot Based On VSLAM

Visual Simultaneous Localization and Mapping(VSLAM)is a technology based on vision sensors for robot positioning and reconstruction of environmental maps.The accuracy and real-time of positioning are crucial.In this thesis,based on the problem of poor positioning and real-time performance of the existing VSLAM system,based on multi-view geometry theory and nonlinear optimization theory,the key algorithms and schemes in the existing VSLAM system are improved and optimized,and theoretical analysis and experiments are carried out through key technical parts.And the overall experiment verified that the improved VSLAM system can be located in real time and accurately on the embedded platform.The classic VSLAM framework is divided into three parts: visual odometer,backend optimization and loop closing.The visual odometer is responsible for local positioning,and the back-end optimization is responsible for global optimization.Loop closing is responsible for identifying the same keyframes and eliminating drift between the same two frames.Based on the current popular RGB-D SLAM system,this thesis uses real-time ORB features for tracking,relocation and closed-loop detection.The visual odometer for the RGB-D SLAM system ignores the area without 3D information,and only uses the 3D corresponding point for camera tracking,which leads to the large drift error problem introduced by long-distance one-way tracking.This thesis proposes a new hybrid corresponding point tracking method.The method fully utilizes the corresponding points of 2D-2D,2D-3D,and 3D-3D to track,and reduces the drift error during the long-distance tracking.The method firstly calculates the 3D world coordinates of the camera's rough pose and 3D-3D corresponding points based on the 3D-3D corresponding points using the IPC algorithm.Then,the 3D world coordinates of the 2D-2D,2D-3D corresponding points are restored using pose and triangulation.Finally,using the BA re-projection method,the pose and 3D world coordinate points are used as initial values,and the optimized camera pose and the 3D world coordinates of all corresponding points are iteratively calculated,and the data sequence collected by the structural sensor is used for experiments.The results verify that the method effectively reduces the drift error.In the optimization part,aiming at the problem of poor global optimization performance of traditional filter algorithm,the beam-adaptive nonlinear optimization algorithm is applied to the back-end optimization,and the Schur elimination method is used to reduce the computational complexity of nonlinear optimization.In the loop closing part based on the word bag model,the similarity score of the a priori similarity normalization is innovatively used to solve the problem of the closed-loop instability of the absolute threshold judgment,and an improved loop closing detection scheme is obtained.In the overall experimental part,the improved visual odometer,back-end optimization and loop closing are combined into a complete VSLAM system,and the real-time and positioning accuracy of the system is verified on the TX2.