Depth estimation from tilted optics blur by using neural network

Abstract

We have been investigating a depth estimation system for real-time usage such as automotive tasks. Conventional method with stereo camera is too sensitive to slight variations of baseline length. Additionally, it has occlusion problem. Conversely, the method that uses a monocular camera by focusing cannot provide a balance between wide-area estimation and real-time estimation. Therefore, we proposed a novel method that adopts tilted lens optics. Herein, our method can obtain depth values at each pixel from the sharpness ratio of only two tilted optics images; our system is consisted of monocular camera system with spectroscopic mirror. Our method uses the optic lens which has some wider angle of view. For that reason, it causes some estimation error based on the difference between the actual camera system and the optical theory. Herein, to reduce the error, we adopted the neural network to calculate the depth value from the blur values and the y-coordinate. In this paper, we report our depth estimation method from tilted optics blur by using neural network.