Camera localization is a fundamental problem in 3D computer vision, referring to estimating the camera pose of an image. As is crucial in applications such as augmented reality, human-computer interaction, visual servo, and robot navigation,it has been extensively studied in the community. Many topics such as 3D tracking of objects, simultaneous localization and mapping(SLAM), image-based localization are addressing the problem in essence. This dissertation investigates camera localization in three different scenarios, i.e. 3D tracking of an object, image-based localization in large scale environments, SLAM relocalization,and focuses on the problems of ease of use, robustness and speed. The main contributions are as follows. An online object reconstruction and 3D tracking system which uses a single webcam is proposed.Unlike the prevalent 3D tracking systems based on prior models,it does not require the users to build 3D models off line. Instead, it automatically reconstructs and tracks the target object after some simple interactions, and hence is more flexible and easier to use.While traditional SLAM techniques are able to reconstruct static scenes online, they can not be used directly for online reconstruction of an object which might be moving.The system goes beyond this and the basic idea is to combine image segmentation and SLAM techniques.On the one hand, image segmentation prevents backgrounds from contaminating the reconstruction process.On the other hand, results of the reconstruction and tracking provide a location priori to improve the accuracy of segmentation.The system also employs some strategies to improve the robustness and stability. An approach for fast localization in large scale environments is proposed.First binary feature instead of real value features like SIFT, which is employed by most existing localization approaches,are used to dramatically reduce the time cost of feature extraction. Besides, a supervised indexing method is proposed to achieve highly efficient approximate nearest neighbor search, which brings about fast 2D-3D matching for localization. The indexing method resorts to randomized trees and uses label information contained in localization databases to train the structures of randomized trees. Specifically, the node tests of the trees are selected to make the numbers of database features in each leaf node as uniform as possible, and matched features collide in the same leaf nodes as much as possible. To...
修改评论