英文摘要 | In 3D computer vision, online real-time visual localization and mapping is important, whose task is to compute a 3D map of environment and spatial positions of visual sensors. As it has wide applications in robot localization and navigation, augmented reality and autonomous driving, it has attracted great attention from researchers. This research direction includes studies on simultaneous localization and mapping(SLAM), online real-time localization and modeling of dynamic objects, and online real-time 6-DOF camera pose tracking from markers. At present, there are still some challenging problems in this direction. For example, the existing SLAM methods cannot achieve high accuracies and fast speed at the same time, the existing monocular SLAM cannot perform well when both camera and objects are moving, and 6-DOF camera pose tracking from markers are not robust and accurate to noise, blur, and long distance. This
thesis studies these problems and the main contributions are as follows.
• A novel stereo visual SLAM framework considering both accuracy and speed at the same time is proposed. The framework makes full use of the advantages of key-feature-based multiple view geometry (MVG) and direct-based formulation. At the front-end, the system performs direct formulation and constant motion model to predict a robust initial pose, reprojects local map to find 3D-2D correspondences by direct formulation and finally refines pose by the reprojection error minimization. This frontend process makes the system faster. At the back-end, structure from motiom(SFM) is used to estimate 3D structure. When a new keyframe is inserted, new mappoints are generated by triangulating. In order to improve the accuracy of the proposed system, a global map is kept by bundle adjustment. Especially, the stereo constraint is performed to optimize the map. This back-end process makes the system more accurate. Experimental results show that compared with the most popular methods ORBSLAM2 and SVO , the proposed framework outperforms them in terms of accuracy and can run at more than 100FPS under CPU.
• An online real-time visual localization and modeling method for dynamic cylinders is proposed. First, images containing a cylindrical object are captured, and the 3D model of the cylindrical object is reconstructed by contour of the cylindrical object and its projective invariance in the images. Secondly, according to the reconstructed 3D model of the cylindrical object, relative 6-DOF camera poses between the camera and the cylindrical object can be tracked online. In the tracking, a linear P3P RANSAC method is proposed to remove outliers. Finally, in order to verify the accuracy of 6-DOF camera poses calculated online in real time, by using the calculated 6-DOF camera poses, a virtual computer 3D model can be projected into the real world and aligned with a cylindrical object in the real world to achieve augmented reality(AR). Experimental results show that compared with the state of the arts, the proposed method achieves higher accuracy and fast speed, and the effect of augmented reality represents the advanced level in the field.
• A class of circular markers is designed and an online real-time 6-DOF camera pose tracking method from them is proposed. We design a class of circular markers. Based on the designed circular markers, the 6-DOF camera pose is analytically expressed as a very precise form by using projective invariance. Afterwards, the pose is further optimized by a novel point-conic bundle adjustment based on a polar-n-direction geometric distance. The proposed method is from imaged circle edges and without PnP, which makes camera pose tracking robust and accurate in terms of image noise, image blur, and distance of camera to the marker. Experimental results show that the localization accuracy of the proposed method outperforms the most popular methods, such as ARToolkitPlus, AprilTag2 and RUNETag, and the localization speed reaches 100FPS under CPU. |
修改评论