Design Of Visual Odometry Based On RGB-D Camera

With the progress of science and technology,robot technology has been widely applied in all walks of life.For autonomous mobile service robots,the robot's perception of the surrounding environment and self positioning in the unknown environment are the basic prerequisites to achieve obstacle avoidance,navigation,environment interaction and other functions.Simultaneous localization and mapping(SLAM)technology can estimate camera's 6-DOF pose through the image sequence captured by camera,and then make robot self localization and mapping.A designed visual odometer system based on RGB-D camera was introduced in this thesis,which is the core part of visual SLAM system.The system is composed of image preprocessing module,feature extraction module,camera pose estimation module and map maintenance module.Firstly,in order to obtain the visual and depth information of the environment,Kinect2 was selected by the image preprocessing module.After calibrated intrinsics and extrinsics of camera,this module obtained the accurate point cloud by using pinhole projection model and image registration.Secondly,according to the requirement of camera localization,ORB feature was selected to design the feature extraction module.In order to ensure the robustness of camera localization,a series of uniform feature point distribution strategies based on the original ORB feature extraction algorithm was proposed,which achieved reliable extraction of image features;Thirdly,according to the requirements of camera localization,different acceleration strategies of feature point matching were analyzed,and the camera pose estimation module based on ICP algorithm was designed to realize the two-stage coarse-to-fine camera pose estimation.Fourthly,considering the real-time performance of the visual odometer,this thesis studied the theory of the sliding window was studied,the jacobian required in the bundle adjustment's optimization was calculated.In the end,map maintenance module implemented the function of mapping and maintenance;Finally,the accuracy test and time-consuming analysis were carried out on the public TUM RGB-D datasets and the actual environment.The experimental results showed that the visual odometer can complete the task of real-time camera pose estimation and sparse map construction.