| Camera pose estimation is an important research topic in computer vision. It is a fundamental and crucial step in the applications of augmented reality. Knowing camera pose, we can edit the content of an image such as generating a virtual object in the image. Currently, many camera pose estimation techniques have been developed focusing on small and indoor environment. For large and outdoor environments, camera pose estimation faces some problems such as poor applicability and high computational complexity.Due to this, in this thesis we develop a two-stage camera position estimation system for large outdoor environments where the SFM (Structure From Motion) and an image retrieval technique are integrated in the system. The main content of this thesis includes the following aspects:Firstly, analyzing the techniques of camera pose estimation and discussing the principle of camera calibration, we develop a two-stage camera position estimation system based on structure from motion and image retrieval techniques.Secondly, in the stage of SFM, the SURF features in each image are extracted and the 3D structures of the scene are reconstructed. After that, a matching table which maps 2D image features to corresponding 3D points is created for subsequent camera pose estimation.Thirdly, in the stage of image retrieval, a method with high retrieval precision, the CNN (convolutional neural network), is employed to extract similar images from the dataset. In addition, the locality sensitive hashing algorithm is also included in the system to improve the retrieval efficiency.Finally, a series of experiments are conducted and results are encouraging. The camera pose of a query image in a large outdoor environment can be accurately estimated which demonstrates the feasibility of developed system. Especially, in the part of image retrieval, the techniques of convolutional neural network and locality sensitive hashing improved the efficiency of the image retrieval. |
PDF Full Text Request