Research On Visual Odometer For Indoor Mobile Robot Based On RGB-D

When a mobile robot completes autonomous navigation in an unknown environment,it first needs to determine own position.that is to realize the positioning function.Laser,inertial navigation system,infrared and WLAN commonly used indoor positioning navigation technology.However,due to the problems of positioning accuracy and cost,these indoor navigation technologies cannot be widely used.Visual odometer is based on the optical positioning technology of camera,and the robot's pose can be estimated by the image feature information.It is cheap and informative.Among them,the vision odometer based on RGB-D camera has gradually become a research focus of domestic and foreign scholars due to its advantages such as fast acquisition speed,high measurement accuracy,and the ability to obtain color information and depth information at the same time.At present,the following problems still exist in the vision odometer based on RGB-D.there are noise and outliers in depth data acquired by camera;When the scene distribution is relatively concentrated,the accuracy of feature extraction and matching is not high,and the accuracy of camera pose transformation matrix is not estimated;The system has a large amount of calculation and slow running speed,which is easy to cause tracking loss.In view of the above problems,this paper makes the following relevant research.Firstly,the imaging model of RGB-D camera is studied.The internal and external parameters of color camera and depth camera were calibrated to generate 3d point cloud.At the same time for RGB-D camera depth in the scene information data the problems of noise and outliers,choose gaussian filtering,median filtering,bilateral filtering experiment contrast,with mean square error,peak to noise ratio and filtering time as evaluation index performance.Bilateral filter with better comprehensive performance is selected for visual odometer.Secondly,for some feature distribution concentration scenarios,when ORB feature extraction is used,the feature points extracted are not uniform,which leads to the problem of large matching error.This paper adopts a form of quadtree to divide features,and completes the extraction and matching of feature points through rough screening and RANSAC algorithm The feature points are uniformly covered in the image,and the image information is fully utilized to reduce mismatching and improve the accuracy of pose estimation.Finally,an improved key frame selection and elimination algorithm is proposed to solve the problems such as single traditional key frame selection algorithm,easy explosion,large amount of system calculation and slow operation speed.The key frames are identified by the motion distance between frames and the local map.The redundant key frames detected by quadratic judgment are deleted.The improved key frame selection algorithm can identify key frames more accurately and timely,reduce the redundancy of key frames,improve the system's operating speed.In this paper,RGB-D vision system and mobile robot are used to build a hardware experimental platform.During the movement of the mobile robot,the scene information is acquired by the RGB-D camera,and the information is transmitted to the upper computer for processing.Experimental results show that the proposed filter can effectively remove the noise and outliers in depth images.ORB feature extraction and matching algorithm can realize feature point homogenization extraction and precision matching.The improved key frame selection and reduce the loss of feature information,and improve the precision and robustness of the robot in the positioning process.