Research On Simultaneous Localization And Mapping Technology Based On ORB Features

In an unknown environment,mobile robots using the sensor carried by itself to Simultaneous Localization and Mapping,SLAM,is considered as the key to realize the autonomy of mobile robots.In the field of augmented reality,unmanned,three-dimensional reconstruction have emerged SLAM figure.With the development of computer vision technology,Vision sensor is widely used in SLAM because of its large amount of data,economy and convenience,monocular SLAM is becoming a hot and difficult research.This paper mainly studies monocular SLAM,the system can be divided into three modules: front-end tracking,back-end optimization and dense reconstruction.The main research work and achievements are as follows:First,research tracking front-end.In order to meet the real-time tracking requirements,choose the ORB features with better performance as landmark.The extraction and matching of the ORB feature points are studied which is the basis for estimation of the pose of the camera.And study PnP algorithm to solve the pose and triangulation method to calculate the 3D position of the features to get a sparse feature points map.For monocular camera scale problems,this paper presents an automatic initialization method based on the improved RANSAC-Eight Point Algorithm,it can automatically select better two images,then estimate the pose between two images by improved RANSAC-Eight Point Algorithm which can eliminate the influence of mismatch and prevent the pose estimation from falling into local optimum,and triangulate a set of initial map points for follow-up.For different datasets,the algorithm can do a good job of initialization,and then the estimated trajectory has global consistency.Secondly,for sparse feature point map can't meet the navigation,obstacle avoidance,3D reconstruction needs,this paper studies dense reconstruction of monocular vision.In this paper,the inverse depth filtering technology is used to achieve the monocular dense reconstruction,its approach is: first,the initial inverse depth estimation of the pixel is obtained by the stereo search constraint,and then by searching epipolar lines of keyframes,get a set of inverse depth estimation of the pixel.This Hypothesis are fused to get inverse model.The system adds a dense reconstruction thread to achieve the above process.In this way,it parallels with tracking and optimizing threads,that will not affect the original operating efficiency to meet the system's realtime requirements,but also completely rebuild the real scene to get a semi-dense map to meet more follow-up needs.Then,the back-end optimization is studied.Introduced graph-based optimization theory,an optimization problem can be described simply and intuitively as a graph,and introduce in detail the entire optimization process and solving graph-based optimization problem.Use it to do bundle adjustment and pose graph optimization,build the corresponding graph and define the error function of edge in graph to optimize the estimated camera pose and the position of features.Finally,the system is evaluated using the TUM standard data set.The results show that tracking rate can be achieved in real time,the estimated trajectory can also meet the accuracy requirements,and get a semi-dense map that can reflect the real scene and has richer information.Experiments on self-collected data sets also has good performance.