An Overview of Monocular Depth Estimation with Applicability in Intelligent Transportation

This paper presents a comparative overview of state-of-the-art monocular depth estimation methods with emphasis on their applicability in intelligent transportation and automated driving systems.

The research evaluates MiDaS, DepthAnything, and ZoeDepth using both relative and metric depth estimation approaches across multiple public datasets including KITTI, NYU Depth v2, and Virtual KITTI 2.

Beyond traditional pixel-level evaluation, the work introduces object-level performance assessment relevant to autonomous driving scenarios, where accurate distance estimation for detected vehicles and vulnerable road users is critical for safe perception and decision-making.

Experimental results demonstrate that DepthAnything achieves stronger generalization performance compared to MiDaS, while fine-tuning ZoeDepth on synthetic datasets improves metric depth estimation capabilities for out-domain environments.

The study also discusses the importance of robustness, computational efficiency, and sensor fusion in deploying monocular depth estimation systems within real-world intelligent transportation applications.