Surf Feature Based Monocular Vision Slam Technology Research And Realization

SLAM is an important issue in mobile robot autonomous navigation and localization under unknown unstructured environment. A considerable progress has been made in SLAM after decades of development. In recent years, with the development of computer vision, monocular visual SLAM adopting only-one visual sensor has gradually become a new hotspot and difficult point in mobile robot SLAM research field.Image feature detection is an important part of monocular vision SLAM. Methods at present can't make a balance between detection speed and feature stability. In order to solve this problem, SURF (Speed Up Robust Features) feature detection operator is adopted to extract image features. This new feature detector not only acquires a comparatively faster detection speed, but also possesses strong robustness to optical, geometric distortion and noise.In order to further enhance the robustness of selected map features as well as improve the localization accuracy, a sub-region based feature significance index random selection method is presented to extract SURF features. The method makes the features have a more uniform, flat distribution in the image.An active vision based elliptic search region fashion combined with the SURF nearest-neighbor matching method is proposed to perform map feature matching. This method greatly reduces the feature search region and accelerates the matching speed.Additionally, Extended Kalman Filter based monocular vision SLAM system structure is further analyzed and EKF-SURF-1pRANSAC algotithm is designed by the combination of the above methods, Inverse Depth Parametrization and 1-point RANdom SAmple Consensus (1pRANSAC). Finally the simulation experiment in unstructured scenes of laboratory is completed by using this algotithm.The experiment shows that SURF based EKF-SURF-1pRANSAC algorithm can accomplish more accurate localization with less map features and has a stronger robustness compared to fast corner based EKF_monoSLAM₁pRANSAC algorithm Moreover, the algorithm has a much faster processing performance in map management and map feature matching modules, and can be used to construct a real-time monocular visual SLAM system.