Depth Estimation using Omnidirectional Stereo Imaging and Machine Learning

In this study, we propose an omnidirectional depth estimation method applicable to autonomous driving systems by combining stereo vision and machine learning techniques. Recently, there has been active research in driver assistance systems and autonomous vehicles, making three-dimensional measurement of distances and surrounding environments indispensable for safe navigation. Traditional sensors like LiDAR and radar offer high accuracy but have limitations in field of view and measurement density. This necessitates the use of multiple sensors, leading to increased costs―a significant challenge for practical implementation.To address these issues, we developed a depth estimation method using Hyper Omni Vision, which enables wide-range and high-density data acquisition. By integrating stereo vision with machine learning, we combine distance measurements from parallax with monocular depth estimation. This approach achieves stable measurements unattainable with conventional methods. Experimental results show that the proposed method reduces detection omissions and improves obstacle visibility. Although it does not yet reach the same level of accuracy as sensors like LiDAR, it has been confirmed that stable and reliable measurements are possible. This indicates its potential as a cost-effective alternative for autonomous driving applications.