Single Camera 3D Visual Localization based on SIFT and Inverse Depth Parameterization Techniques

• Main Authors: Ko -Ting Yeh, 葉科廷
• Other Authors: 陳傳生
• Format: Others
• Language: zh-TW
• Published: 2007
• Online Access: http://ndltd.ncl.edu.tw/handle/06540994543616362177

碩士 === 元智大學 === 機械工程學系 === 95 === When an autonomous mobile robot operates in an uncertain environment, one of the fundamental tasks is self-localization. The localization task is associated with map building, and there is SLAM (Simultaneous Localization and Mapping) technical development afterward. To solve this problem, the vision-based SLAM algorithm develops rapidly in recent years. The method, called vSLAM, utilizes some features of the object in image to provide localization information of the camera. The features of the object that vSLAM means are in the images with the strong contrast in local image(e.g., edge and corner). But affine transform caused by distance changes and rotation of the object will get the wrong information of the features. This paper uses the theory of SIFT (Scale Invariant Feature Transform) to detect the features and improves the shortcoming of above-mentioned methods, and adopts inverse depth parameterization theory to solve the expression of the features in the infinite. SIFT theory is besides being able to filter the features on the edge, and the description for the features also include some local images. For this reason, it produces the particular descriptor. We utilize the descriptor uniqueness in difference images to match the features and learn the movement situation of the features in continuous images . However, the detected features in all scale with SIFT cannot be totally suitable for vSLAM, which change heavy and have low matching degrees would reduce the accuracy of localization. For this problem, we utilize the concept of Large Scale to select the features with high matching degrees and robustness, and realize the single camera 3D visual localization based on EKF.