Research And Implementation Of Monocular Vision SLAM System

With the continuous development of science and technology,SLAM technology has become one of the key technologies for robots to become autonomous and intelligent.At the same time,with the application of driverless technology,SLAM technology has also attracted more researchers.However,the current visual SLAM algorithm still has the problem of poor accuracy,which affects the operation of the system.In the SLAM algorithm based on feature point matching,the accuracy problem mainly exists in feature extraction and matching and depth estimation optimization.In order to solve the above problems,the research work of this paper includes the following three parts:1)An image feature point matching algorithm based on density-ORB feature is proposed.Firstly,we studied the mainstream feature extraction algorithm and compared the feature extraction results,and found that the ORB algorithm is significantly better than the SIFT and SURF algorithms in performance.Then,the Steer BRIEF descriptor in the ORB algorithm only determines the 0/1 code by comparing the gray information of two pixels,which is easy to generate the feature point mismatch phenomenon,so we proposed a density-ORB feature matching algorithm based on pixel density.The experimental results on the Oxford University Standard Library showed that the algorithm has better experimental performance in feature extraction.2)The depth filter model is integrated into the back-end optimization algorithm.Firstly,we constructed a Gaussian pyramid on the image,and matched it in the low-resolution image to limit the search range of the high-resolution image.Then,we used Patch Match technology in each layer of image to generate a depth value for each pixel,and used the similarity of adjacent pixel depth values to transmit reliable depth values to surrounding pixels.Then,the depth filter of the uniform-Gaussian mixture distribution is used to estimate the probability distribution of the depth information of the map point until convergence,and finally the coordinates of the three-dimensional space point are calculated by the back projection function.The experimental results on the standard data set REMODE showed that the algorithm can obtain reasonable and accurate depth estimation results,which has certain superiority and robustness.3)A ROS-based monocular vision SLAM system is implemented.In order to integrate the monocular visual SLAM algorithm into practical applications,according to the basic theory of visual SLAM,a ROS-based monocular vision SLAM system is realized.The main purpose of the system is to draw camera trajectories in real time and generate sparse 3D map.The results of testing the system in the real scene of the laboratory showed that the system can calculate the trajectory of the camera in real time and generate a sparse 3D point cloud map.