The Research Of Simultaneous Localization And Mapping Based On RGB-D Images

Recently,simultaneously localization and mapping(SLAM)is a popular research problem,and it is a key issue for intelligent robots to realize autonomous navigation in unknown environment.SLAM mainly depends on sensors for acquiring information,to realize the functions of building environment map and real-time positioning at the same time.The RGB-D camera has many advantages.For example,it can capture both the color images and the depth images at the same time.The speed of acquiring information is fast,and its measurement accurancy is high.Also its price is very low.Thus,the RGB-D SLAM has gradually become a hot research issue among scholars at home and abroad.The steps of RGB-D SLAM method are as follows:firstly,the features are extracted and matched according to the color images,secondly,the original transformation matrix of the camera is estimated using feature matching point pairs combining the depth information;thirdly,the pose transformation matrix of the camera is estimated accurately;finally,the pose figure is optimized,and the camera pose is corrected through the closed-loop test.There are some problems in RGB-D SLAM at present,for example,RGB-D data contains the noise and redundancy,the estimated accurancy of the pose transformation matrix is not high,the feature matching accurancy is not high;the computational cost of feature extraction is high.To solve above problems,three works are investigated in the thesis,and the details are as follows:(1)For the problem of RGB-D camera data containing noise and redundancy,the point cloud filter method is improved in the thesis.For the estimated accuracy of the pose transformation matrix is not high,the ICP(Iteratively Closet Point)algorithm is used to optimize the camera pose and improve the estimated accuracy of the pose transformation matrix.At the same time,the criterion which determines whether to use the matrix of pose optimization algorithm as the pose transformation matrix of camera is based on the degree of convergence of the ICP algorithm.The point cloud filter method and the pose optimization algorithm are used to improve RGB-D SLAM method.The improved RGB-D SLAM method can effectively improve the accuracy of RGB-D SLAM under the premise of without increasing the computational cost.(2)In view of the problem of the low feature matching accuracy in the RGB-D SLAM,the thesis introduces the depth information to improve the feature matching algorithm.At the same time,the pose optimization algorithm which has higher accuracy is used for the RGB-D SLAM method,and it is named NDT(Normal Distribution Transformation)algorithm.The RGB-D method combining the above two algorithms improves the accuracy of RGB-D SLAM.(3)The thesis uses a higher real-time feature extraction algorithm which is named AKAZE(Accerated-KAZE)algorithm,and uses the improved point cloud filter method to solve original point cloud.The RGB-D SLAM method improves the real-time performance of the RGB-D SLAM system under the condition of ensuring the positioning accuracy.The thesis verifies the performance of these three RGB-D SLAM methods on the international public datasets,and gives the experimental results about feature extraction and matching,point cloud filter,point cloud splicing.The three RGB-D SLAM methods are evaluated in terms of aspects of real time and accuracy,and the experimental results are analyzed.In the end,the thesis summarizes the research works,and proposes some problems remained to be researched.