Radar-Camera Fusion Network for Depth Estimation in Structured Driving Scenes.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Depth estimation is crucial for autonomous driving perception systems.
• Existing methods often neglect latent semantic information in depth map reconstruction.
• Fusing multi-modal sensor data like radar and RGB images is a promising direction.

Purpose of the Study:

• To propose a novel dual-branch network for dense depth map prediction using fused radar and RGB images.
• To leverage latent semantic information in driving scenes for improved depth estimation.
• To enhance the focus on critical driving areas during training.

Main Methods:

• A dual-branch network architecture is proposed, dividing the driving scene into three parts for depth prediction.
• A fusion strategy is implemented to merge depth maps from different scene parts.
• A variant L1 loss function is utilized to prioritize areas of interest during training.
• The method is evaluated on the nuScenes dataset.

Main Results:

• The proposed method effectively predicts dense depth maps by fusing radar and RGB data.
• The dual-branch network successfully utilizes semantic information for improved accuracy.
• Experimental results on the nuScenes dataset show superior performance compared to state-of-the-art methods.
• The variant L1 loss function aids in focusing on important driving regions.

Conclusions:

• The proposed dual-branch network offers an effective approach for depth estimation in autonomous driving.
• Fusing radar and RGB data, combined with semantic awareness, significantly improves depth map accuracy.
• The method demonstrates a promising direction for enhancing autonomous vehicle perception systems.