Research Of Visual Odometer Based On RGB-D Camera

As the basis of automatic navigation and positioning of mobile robot,vision odometer has been a research hotspot in the field of robot vision technology since the concept of SLAM was proposed.At present,RGB-D camera can directly obtain the pixel depth information of the image,avoiding the depth calculation process of 2D-3D transformation,so the research of visual odometer based on RGB-D camera plays an important role in SLAM,and it is also a hotspot at home and abroad.However,there are still some problems in the visual odometer based on RGB-D camera,such as how to deal with the inevitable noise;feature point extraction and matching accuracy affect the calculation accuracy of camera pose;the large consumption of resources in calculation,slow running speed and data loss.In view of the above problems,this paper makes the following related research:1.Through the mathematical representation of rotation and translation of rigid body in three-dimensional space,the rotation matrix and transformation matrix were calculated by using the eigen library,and the expression and transformation of pose were described by using the Lie group SO(3)and SE(3)and the corresponding Lie algebra so(3)and se(3);the derivation of Lie algebra was carried out for the disturbance of Lie group through the linear approximation of BCH,for reducing the error;2.Using the distortion function of Open CV to process the image;processing the depth image of RGB-D camera to form a three-dimensional point cloud map;fitting the same curve by three nonlinear optimization methods,Gauss Newton method,cere and G 2O library,and making comparative analysis;3.Through the calculation and matching of fast key points and BREF descriptors of orb feature points and the Hamming distance screening of matching points,the matching condition of orb feature points of two adjacent was gotten.4.Build a visual odometer for the software system through Linux,and obtain the RGBD test set of TUM standard as the source of data;then verify the indoor environment,large indoor environment,texture environment,and environment with dynamic objects separately,and finally output the camera Movement track.In this paper,the visual odometer based on RGB-D camera is realized by processing the scene information obtained by processing RGB-D video data in Linux system.The results show that: Gauss-Newton method,Cere and g2 o methods all can achieve better curve fitting,but the calculation efficiency is significantly different;ORB feature point extraction and matching algorithm can well achieve the feature point uniform extraction and matching;After verification in four environments,the visual odometry results achieved are robust,but algorithmic optimization is still needed in the texture environment and the environment,to improve the accuracy of its motion trajectory estimation.