A geometry-based 3D reconstruction from a single omnidirectional image

Abstract

This paper proposes a method 3D scene reconstruction from a single omnidirectional image based on geometry recovery in 3D coordinate system. Different with previous work which only deals on scale model, the proposed method can obtain real size of 3D scene which required for real application such as robot navigation. The proposed method is divided into two mainly parts, region classification and reconstruction. In the first step, region classification using ground detection is applied to get pixel whether belongs to either ground region or vertical surface region. Once the ground pixels are determined, the remaining pixels will be classified as vertical surface. All pixels on the ground region will be reconstructed by ground surface reconstruction model, meanwhile all pixels on the vertical surface region will be reconstructed by vertical surface reconstruction model. The experiment shows that the proposed method has some advantages such as simple process, low-cost computation and precise real coordinate system. Hence, the proposed method is expected to be implemented for real application.