Research On Monocular Visual Odometry For Mobile Robot

With the rapid development of many disciplines such as artificial intelligence,mobile robots are widely used in many fields such as industry and civil.Positioning is the key and foundation for intelligent and autonomous operation of mobile robots.The traditional odometer uses a wheel odometer,a laser measuring instrument,an infrared ranging sensor,etc.to realize the positioning of the mobile robot,so that there are problems such as wheel slippage,large measurement error,and large terrain factors.The visual odometer can realize the positioning of the mobile robot only by acquiring the image information,and can be applied in the case of GPS failure,and has become the first choice for the mobile robot to achieve positioning in recent years.The monocular visual odometer has the advantages of simple structure,low cost and high calculation efficiency.In order to make the mobile robot operate stably in the indoor environment,the image acquired by the monocular vision sensor is used as the research data,and the monocular visual odometer system is realized by the feature point method.The main results of this paper are as follows: First of all,this paper analyzes and compares the current mainstream feature detection algorithms such as SIFT,ORB and SURF.Finally,the ORB detection algorithm is selected as the feature point detection algorithm in this paper,and for the disadvantages of uneven distribution of ORB feature points in the image,region segmentation of the image achieves ORB feature point homogenization.Secondly,for the case of mismatching in the image feature matching process,the RANSAC algorithm is used for mismatching rejection;By simultaneously calculating the Homography Matrix for the plane and the Fundamental Matrix for the non-planar,the appropriate model was selected to realize the automatic initialization of the monocular,and the initial pose of the camera and the point cloud map were obtained.Then,using 3D-2D motion estimation to solve the initial pose of the camera,as the incremental motion error of the camera continues to accumulate,the pose is constructed for the camera pose,and through the G2 O library to achieve nonlinear optimization of camera pose;At the same time,in order to improve the efficiency of system operation,this paper proposes a suitable key frame selection strategy,which can effectively avoid the impact of redundant data on system operation efficiency.Finally,in order to verify the actual running performance of the visual odometer,in the indoor environment,experiments such as linear motion,curved motion,and closed loop motion were performed.And compare with the total station measurement point data,the experimental results show that the relative error and cumulative drift error of the system are within 5%.The experiment verifies that the accuracy and robustness of the visual odometer meet the positioning requirements of mobile robots.It provides the basis for navigation obstacle avoidance and path planning of mobile robots.