A Monocular Depth Estimation Method for Autonomous Driving Vehicles Based on Gaussian Neural Radiance Fields.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Monocular depth estimation is crucial for autonomous driving, providing scene depth from single images.
• Current methods struggle with visual artifacts, scale ambiguity, and occlusion, limiting performance in complex environments.

Purpose of the Study:

• To develop an improved monocular depth estimation method for autonomous driving using Neural Radiance Fields (NeRF).
• To address limitations of existing approaches, including visual artifacts, scale ambiguity, and occlusion handling.

Main Methods:

• Introduced a Gaussian probability-based ray sampling strategy to manage large, complex scenes and reduce computational load.
• Designed a lightweight spherical network with fine-grained adaptive channel attention for detailed feature extraction.
• Mapped pixel-level features to 3D spatial locations for enhanced NeRF model generalizability.

Main Results:

• The proposed NeRF-based method demonstrated superior performance on the KITTI benchmark for depth estimation tasks.
• Achieved significant improvements over traditional monocular depth estimation techniques.
• Showcased enhanced model efficiency and generalization capabilities.

Conclusions:

• The developed method offers significant advancements for practical monocular depth estimation in autonomous driving.
• The approach effectively tackles challenges in complex environments, improving reliability and applicability.
• This work paves the way for more robust perception systems in autonomous vehicles.