Study On Monocular SLAM Based On Graph Optimization In Natural Environment

Simultaneous Localization and Mapping (SLAM) is a process of robots to complete the location and map building according to their self-position estimation and sensor data. SLAM is the foundation and key technology to achieve autonomous navigation and it has become one of the hot topics in other research fields such as AR, visual medical,3D reconstruction, intelligent household. Because of the visual sensor accessing large amount of information, wide application, convenient installation, inexpensive and many other advantages, visual SLAM has become a hot issue of researches at home and abroad. To describe the natural environment characteristics rapidly and stably, build the indoor and outdoor large-scale map effectively and optimize the data associated with the system are the difficult points in the SLAM field and the focus of this article. Monocular SLAM algorithm modules of this article can be summarized as follows:natural landmark feature extraction, Key-Frames detection, loop closing detection; motion structure recovery, map optimization and inverse depth estimation. The main work and innovations are as follows:Firstly, the influence of different image features algorithm and features distribution on the performance of SLAM is studied. Considering of the requirements of both image feature extraction speed and stability, we select the ORB binary feature extraction algorithm to describe the natural environment. To enhance the robustness of the feature map and improve the positioning accuracy, an optimization ORB feature extraction method based on region segmentation was proposed. In order to achieve large-scale environment construction, Graph Optimization model is processed by Key-Frames only. Technique fused uniform sampling in time and image matching was taken as the method of choosing Key-Frames by comparing the principles and the pros and cons of various scenarios sampling.Secondly, on the basis of the image feature algorithms and Key-Frames, the bag-of-words model with a growing vocabulary tree is built to detecting loop closing for optimizing data association. The time continuity and spatial consistency are used as constraints to improve the correctness of loop closing detection, and the loop closing scheme is applied to the case of special tracking lost.Then, the camera imaging model is structured to complete the calibration of camera and structure from motion is restored by using singular value decomposition (SVD) in the essential matrix of Key-Frames. Pose graph based on the data between different time pose is established and optimized by using LM least squares. Using inverse depth filtering model, the depth of the map is obtained. Finally, the feasibility of the proposed monocular SLAM algorithm is verified with experiment in both the indoor and outdoor natural environment.