Passive Absolute Distance Estimation for Maritime Navigation
Jon Muhovič, Janez Perš · 2024
Accurate distance measurements are crucial for safe navigation of unmanned surface vehicles, in order to detect potential obstacles and avoid collisions. However, accurate depth sensors can be prohibitively expensive, heavy, or power-hungry. In this work, we explore image-based alternatives for estimating distances to observed scenes and compare their predictions with absolute measurements obtained from an onboard LIDAR sensor. Several stereo reconstruction methods and several single-view depth estimation methods are evaluated and their accuracy in different distance ranges is reported. We show that state-of-the-art single-view methods such as Metric3D and MiDaS can perform well, even if no domain-specific fine-tuning is applied, showing the potential for their use onboard real vehicles.