Paper Title
Omnidirectional Monocular Distance Estimation Applicable to Autonomous Driving

Abstract
This study proposes a 360° monocular distance estimation method using machine learning that can be applied to autonomous driving. With the increasing development of autonomous driving technology in recent years, the private sector has seen a rise in its usage, making it a familiar technology.In this context, sensors serve as the "eyes" of a machine and play a crucial role in achieving higher performance at a lower cost.This study primarily focuses on depth measurement through cameras, which offer a wide area and high density coverage.Depth estimation using cameras has been studied for a significant period of time.Cameras are one of the essential sensors for recognizing the surrounding environment.Moreover, the development of machine learning technology in recent years has significantly improved image recognition capabilities, expanding the range of applications for cameras. Based on the above, we propose an omnidirectional monocular depth estimation method involves the use ofour omnidirectional visual sensor and machine learning. our omnidirectional visual sensor is a device that combines a hyperbolic mirror and a camera. By placing the center of the camera lens at the focus of the hyperbolic mirror and taking a picture, an omnidirectional image can be captured. This image can be converted into a panoramic image using image processing. Using this characteristic, our method captures an omnidirectional image through the sensor and obtains a panoramic image through image processing. Then, depth estimation in all directions is performed by monocular depth estimation using machine learning. However, the distance value obtained by monocular depth estimation using machine learning is a relative value, so the actual distance could not be obtained. To overcome this limitation, our study proposes a calibration method that estimates the absolute distance using regression curves.To validate our proposed method, we created a prototype device and conducted three experiments, namely a depth image generation experiment, a distance measurement experiment, and an absolute distance image generation experiment. In the depth image generation experiment, the proposed method was able to generate accurate depth images with fewer detection omissions and obstacle visibility when compared to the conventional method. In the distance measurement experiment, we demonstrated that our proposed method can measure distances with an accuracy approaching that of the conventional method, even with a single eye. The standard deviation was up to 29.1 cm lower and the resolution was up to 31.0cm higher than that of the conventional method, indicating the possibility of highly stable measurement. In the absolute distance image generation experiment, we showed that the use of regression curves can enable absolute distance estimation, overcoming the limitation of depth estimation using machine learning based on relative values.As a result, the proposed method can provide accurate and reliable measurements applicable to autonomous driving. Keywords - Monocular Depth Estimation, Omnidirectional Camera, Distance Estimation, Autonomous Driving Omnidirectional