Study On Monocular Visual Simultaneous Localization And Mapping Under Large-Scale Environment

Simultaneous Localization and Mapping(SLAM),through a sensor's perception information of an unknown environment,estimates the motion trajectory of the sensor in the environment,and builds the structural map of the environment at the same time.The visual SLAM algorithm through a monocular camera is one of the hotspots in the field of visual SLAM.There are still many difficulties and challenges about how to accurately and robustly estimate the real-time and long-term trajectory of the camera,and to build the environmental structure map simultaneously.In this thesis,a monocular visual SLAM algorithm in a large scale environment is implemented with good real-time and robust performance.The thesis focuses on two modules in the SLAM algorithm,the visual front-end module and the visual back-end optimization module.The visual front-end module mainly completes the feature detection and correlation,the camera pose estimation and the initial calculation of the map point.The back-end optimization module is mainly completed keyframe topology maintenance,camera pose and map point optimization.The main works and innovations are as follows:Firstly,an adaptive uniform binary feature detection algorithm is proposed.The algorithm adjusts the threshold of the feature points adaptively according to the contrast of the image,so that the feature points detected by the various types of images are kept as stable as possible.While the detected features were organized in the quadtree form,a specific number of features were randomly selected in a combination way of depth-first search and breadth-first search,so that the feature points were randomly distributed evenly to the entire image,thus reducing the dynamic target characteristics and effects of some abnormal value on the camera pose calculation accuracy.Compared with the classical ORB feature detection algorithm,the algorithm proposed in the thesis is more stable in the estimation process of the camera trajectory,and the accuracy of the trajectory estimation is higher and the real-time performance is also kept on the same order.Secondly,a visual front-end algorithm based on adaptive uniform binary feature is proposed.The algorithm completes the feature detection,feature search matching and data association on adjacent video frames;and uses the normalized eight-point algorithm based on RANSAC and the algebraic outlier decision based on the EPn P algorithm,calculates the relative pose matrix between the two frames,reconstructs the map points corresponding to the feature points by triangulation.The average run time of the entire visual front-end module in the KITTI Odometry data set and the self-built Xidian University North campus dataset was respectively about 30 milliseconds and 50 milliseconds,has good local precision,robustness and real-time performance.Finally,a back-end optimization algorithm based on the FAB-MAP is implemented.The back-end optimization module is divided into two parts,the similarity detection section and the map optimization section.The similarity detection part is based on the FAB-MAP algorithm,which completed the loop detection and the maintenance of the keyframe topology graph,including the clipping of the keyframe,the cropping of the map point and the updating of the topology graph.The map optimization section uses the g2 o to complete the three optimization strategies of local bundle optimization(Bundle Adjustment,BA),random sparse global bundle optimization and loop closure optimization on the key frame topology.The thesis implements a monocular vision SLAM algorithm which can calculate the camera movement trajectory of all the test data set.It has good robust performance and global precision performance,and the algorithm could reach the real-time performance with an average running time of 80 milliseconds per frame on the experimental platform of the thesis.