Improved RGB-D Visual Odometer Based On Kinect

The research content of this project comes from the inspection project of bridge UAV Based on multi-source seamless positioning and navigation of Shanghai Science and Technology Commission.With the rapid development of artificial intelligence,mobile robot is gradually in an important position.The main navigation is one of the important symbols of the intelligence of the mobile robot.In the process of autonomous navigation,the robot first needs to determine its own position,that is,to realize the real-time positioning function..The positioning and navigation of mobile robot is an indispensable link to realize intelligence.Visual SLAM(SLAM)has been committed to solving the positioning,drawing and navigation functions of mobile robots in unknown environments.While moving,mobile robots detect the surrounding environment information and build maps to determine which direction to go.Location and Mapping of Mobile Robot are the two most important parts of Visual SLAM.But the vision odometer also has problems as it is for rgb-d cameras for color and depth image capture,in the actual detection,the detected depth image has a plurality of areas with missing depth information due to illumination,occlusion,speckle and the like,and the feature points of the missing depth information at the time of the serious detection reach a quarter of the total number,There is no doubt that the initial pose estimation and the later iterative optimization result in a great error;secondly,the current mainstream ORB algorithm does not have a good solution at the scale and the illumination problem of the feature point,leading to the poor robustness of the odometer in the light and scale problems.In view of the above problems,the improved RGB-D vision odometer based on Kinect,which is based on Kinect,is designed to ensure the visual odometer On the basis of real-time and robustness,the main contributions of visual odometer are as follows:(1)Several mainstream feature point methods in visual odometer are studied,and the current mainstream feature point method ORB algorithm is improved.On the basis of ensuring the good real-time and rotation invariance of the ORB algorithm,referring to the idea of the BRISK algorithm for the scale problem,after establishing the image pyramid toobtain the position and scale of the feature,the average sampling mode is used to construct the scale.The descriptors normalized with lighting,because the descriptor itself uses a symmetrical circular sampling structure,which makes it adaptable in rotation transformation.In addition,on the scale problem,because the location and radius of the sampling point change with the transformation of the scale,it also has good adaptability on the scale.In terms of illumination invariance,an adaptive threshold based on image contrast is proposed,so that the improved algorithm will not lose a large number of feature points when the contrast of the image is greatly reduced,which makes the ORB algorithm better in the lighting problem.Robustness,and subsequent tests also proved the feasibility of the algorithm.(2)Regarding the positioning accuracy of the odometer,it was found that in the actual detection,the depth camera often lost the depth area due to various conditions such as lighting,occlusion,speckles,reflections,and absorption.This makes it difficult to achieve good results in iterative optimization of camera pose estimation using traditional ICP and other algorithms,and even makes the initial estimated pose deviate from the iteration interval,making iterative algorithms difficult to successfully converge.First,the detected feature points are removed from the outer points using the RANSAC algorithm,leaving reliable inner points.The 3D-3D point pair with good depth information uses the ICP algorithm to estimate the camera's initial pose,and the 3D-2D point pair missing from the depth camera uses the Pn P algorithm to estimate the camera pose.The camera pose is weighted to obtain the total camera initial pose.It is estimated that the point-pair problem is turned into a BA model,the cost function is obtained and iterative optimization is performed by using the g2 o solver.Finally,the experiment also proves that the algorithm has a good practical effect.On the other hand,it also improves the robustness of the visual odometer in harsh environments.