Research On Key Technologies Of Ship Surface Transportation Based On Monocular Visual SLAM

The transportation ship surface work is the basic work to ensure the daily training and combat of warships.As the number and variety of aircraft increases,the deck area increases,The traditional Artificial transportation method is time-consuming and laborious,east to be influence by subjective,and is prone to accidents when the deck is busy.Visual Localization is a new method that is increasingly admired by researchers in the continuous development of computer technology.Compared with traditional Localization methods,it has a high degree of autonomy,is not easy to interfere,and is cheap and can guarantee high precision.These characteristics make it a well application in transportation work of ship surface.This paper proposes a solution based on isual Simultaneous Localization and Mapping(Visual-SLAM)algorithm to solve the problem of aircraft localization during the process of transportation.The front end of the visual SLAM algorithm is also commonly called Visual Odometry,It’s camera motion is estimated based on the adjacent image information.The SLAM algorithm based on the feature point method has high requirements for inter-frame pose estimation during initialization and tracking.In the initialization process,usually uses the 5-point method to decompose the essential matrix,and the8-point method decomposes the homography matrix.Based on the existing algorithm,Our paper combines the engineering application reality: the transportation work can be regarded as the planar motion and use the camera installation angle as constraints,Reduce the 6-degree-of-freedom problem of traditional Visual SLAM to 3 degrees of freedom by constrains,Improved method for decomposing homography matrix,Finally more matching inliers are obtained,and improves the robustness of the initialization algorithm in the planar scene.The basis of the solve of the relative pose between frames is the matched feature points.The matching of the feature points usually end with the descriptor and the RANSAC algorithm to get inliers.The complexity of the RANSAC algorithm increases exponentially with the number of matched points brought into the calculation.Based on the improvement of the accuracy of inter-frame pose estimation and the reduction of RANSAC culling outliers,we propose that when the camera is forward-looking and follow planar motion is perpendicular to the direction of gravity,other methods can be used to correct the direction of gravity during motion.The rotation matrix only has a rotation angle around the Y-axis.At the same time,the far and near points in the feature points are separately calculated,while the far point is sensitive to rotation,the near point is sensitive to the amount of translation.The inter-frame pose can be estimated using only 1.5 pair points.After replacing the algorithm in the open source ORB-SLAM algorithm,the effect is better than the original algorithm on multiple data sequences of KITTI,which proves the feasibility and robustness of the algorithm.The traditional point feature based Visual SLAM algorithm only uses the image coordinates of the feature points,while the number of matching points will decrease significantly under the influence of large-angle affine distortion.The ASIFT feature points increase the feature point local affine transformation based on SIFT,which not only overcomes the large angle tilt,but also provides an affine relationship for feature matching.In the case of planar scene or vertical direction alignment,the number of inliers of the decomposition homography or the essential matrix can be reduced,and the complexity of RANSAC iterations is reduced.The results verified by real data sets show the feasibility of the algorithm,In the ORB-SLAM framework,the features and inter-frame pose estimation are replaced to form ASIFT-SLAM,The performance of ASIFT-SLAM on the KITTI dataset proves to be better than ORB-SLAM in accuracy.