| Along with the development of science and technology, Autonomous Mobile Robot is a comprehensive system integrating the functions of environmental perception, decision-making and control and implementation. It has been gradually permeated into most fields of human activities. Precise positioning and environment reconstruction are the prerequisite conditions of autonomous navigation for Autonomous Mobile Robot. Relying on its many advantages such as low cost and large amount of information, Simultaneous Localization and Mapping based on visual has gradually become the important developing direction in the SLAM field.Based on the unmanned vehicle driving platform of The Institute of Applied Technology, Hefei Institutes of Physical Science, C.A.S, the unmanned vehicle’s Simultaneous Localization and Mapping under the outdoor environment which is dependent on Bumblebee2 binocular stereo vision camera is researched in this dissertation. Firstly, the basic principle of image processing are introduced. Secondly, a strong emphasis on the study of stereo image matching and tracking, Visual SLAM modeling and a Visual SLAM model based on the Graph Optimization has been laid in this dissertation. Finally, on the basis of the Levenberg-Marquardt algorithm, the optimization objective function on the position is established. The visual position, map construction and optimization problems based on Levenberg-Marquardt algorithm of unmanned vehicle model are completed in this dissertation.Specific to many problems such as the poor feature points, obviously illumination change and big movement between two image frames under outdoor environment, an algorithm named simultaneous matching and tracking is proposed in this dissertation. For the feature extraction, the Gaussian filter is used for image preprocessing, and after the comparative analysis of Harris, SIFT and SURF feature extraction from the aspect of operator theory and experiment performance, the Harris corner is finally chosen as the base of detection process and feature descriptor construction. For the image matching, with the sum of absolute difference as the matching criteria, the image is partitioned for the purposed of equalization distribution. What’s more, the RANSAC algorithm is introduced to improve the precision. Finally, the closed-loop constraint is used for the matching and real-time tracking, and the closed-loop confirm system is adopted to optimize the precise. Then, compared with SIFT and SURF in matching and optical flow in tracing, the superiority-inferiority on visual SLAM application of our algorithm is stated.The Levenberg-Marquardt nonlinear least squares algorithm is adopted to build optimization-based visual posture model and Map model of the unmanned vehicle. Moreover, the Graph Optimization is used for eliminating the cumulative error. For the position of the unmanned vehicle, the sparse three-dimensional point is firstly rebuilt according to the set of matching and tracking, and then the error optimization function between observation and forecast of 2D-3D is built based on the LM algorithm. Then, the rotation matrix R and translation matrix T can be solved by obtaining the optimum solution of a nonlinear multimodal optimization function. Finally, by adopting Multi-start global search, the best R/T matrix is solved and accurate positioning is given out. For the map constructing, the LM algorithm is used to build the optimization equation on observation signpost, and the Graph Optimization strategy is adopted to establish the across data correlation for the purpose of optimizing the Mapping error. Finally, a real-time graph is established according to the simplified topological graph. What’s more, according to the disparity map and three-dimensional point set, the local dense map and sparse global map of point cloud are given out, respectively. Finally, experiments in three different scenes are designed, and results show that both the accuracy and rapidity of visual SLAM are identical with the expectation. |
PDF Full Text Request