A Depth Estimation Method Based on Catadioptric Imaging System Using an Omnidirectional Camera and a Spherical Mirror

Abstract

This paper proposes a new method for estimating the depth of an observation space from a monocular image based on a catadioptric imaging system using an omnidirectional camera and a spherical mirror. This method uses a catadioptric imaging system with a curved mirror, which enables observation of a large space at a time, and uses only one omnidirectional camera as the sensor of the imaging system. Therefore, the method does not depend on learning and can acquire a wide range of 3D information with a compact device. Since 3D information is useful for 3D scene recognition, the technologies for acquiring and displaying 3D information are rapidly spreading in various fields. Laser light and multi-view images are known as sensing methods for 3D information, and estimation technologies based on deep learning continue to develop. However, each of these methods requires large-scale equipment or observation systems, or prior information. We compare our method with other methods and confirm its effectiveness through demonstration experiments using images taken in an environment constructed using CG.