특징점의 구 사상을 활용한 전방향 영상으로부터의 3 차원 영상 지도 생성 기법

Abstract

In this paper, we propose a robust and fast 3D data reconstruction method for large FoV(Field-of-View) cameras. Large FoV cameras are beneficial in many vision systems because they increase visual overlap and capture more pixels of data from the environment. However, most vision systems operate based on the pinhole camera model. This leads to problems when using large FoV cameras, such as distortion or decreased resolution. Other methods address this problem by using different camera models (e.g., mapping overall images onto a cube surface), but this increases the amount of data processing required, and distortion is not completely eliminated. To address these issues, we propose a method in which 3D data is generated from large FoV images using sphere mapping of features. The proposed method has two advantages. The first is robustness since the sphere model uses the full FoV without distorting the images. The second is speed, arising from the fact that only the feature points are mapped and not the entire image. The proposed system is compared with others on various image sequences. The results show that, within the same amount of processing time, our system can generate a greater number of successful mappings, with greater accuracy, compared to the other approaches.