Research On The Application Of Stereo Vision SLAM In Mobile Robot

Driven by the development of science and technology,robots come into our lives.Whatever the purpose of the robot,it needs to have intelligent perception of the surrounding environment and know its location in real time as long as the robot needs to move autonomously.At the same time,localization and mapping(SLAM)technology came into being.Visual SLAM technology can carry different sensors,and binocular cameras are the most widely used.According to the different treatments of the images,the present study is based on feature keypoints.It lies in real-time,accuracy,and robustness.Aiming at these core problems,this thesis based on the ORB-SLAM2 algorithm,using stereo camera to study the Visual SLAM system,the main research contents including the following aspects:Overall study and research on Visual SLAM:Study of the principle of the stereo camera imaging model,and grasp the camera coordinate transformation,the learning and application of stereo camera calibration to obtain the internal parameters and the image distortion correction,the content of every part of the ORB-SLAM2 algorithm in-depth study,the algorithm ported to mobile robot in practical application,build sparse point cloud map for the environment,and to eliminate cumulative error using the loop detection function.In view of the algorithm to locate the problem of low accuracy,and combining with the actual application of image rotation problem,this thesis puts forward the improved ORB feature descriptor arrangement,under the different environment to test,the result shows that the improved algorithm improves the accuracy of feature matching by 33.72%,positioning accuracy improved by 0.09%.Aiming at the question of movement speed of the algorithm can not meet the real-time requirements,this thesis puts forward a kind of acceleration method based on CUDA(Compute Unified Device Architecture,Architecture Unified computing equipment),it is mainly used for processing algorithm of time-consuming partfeatures extraction and matching part,made full use of GPU's high computing performance,experimental results show that the speed of feature extraction and feature matching is improved by 65.3%and 86.4%respectively.Taking full advantage of the vision odometer and the idea of orb-slam algorithm,this thesis proposes a calibration method based on the external parameters between the cameras without common vision,which solves the problem of the initial position relation between the cameras,and is convenient for the subsequent multi-camera system to get a consistent environment map,and experimental results show that the calibration method is simple and feasible.On the white wall or grain rational poor areas,the camera can not extracts the feature keypoints leading to localization failure,this thesis proposed multiple stereo cameras system based ORB-SLAM2.This problem is solved by placing multiple stereo cameras at different angles to obtain environmental information of a wider field of view.When one camera faces the white wall,the other camera can extract the environment features and continue to work.At the same time,due to the more abundant information obtained by multiple cameras,the effect of building environment by mobile robot is better and the positioning accuracy is improved.The experimental results show that the positioning accuracy of the multi-stereo system is improved.