SemanticDepth: Fusing Semantic Segmentation and Monocular Depth Estimation for Enabling Autonomous Driving in Roads...

Area of Science:

• Computer Vision
• Autonomous Driving Systems

Background:

• Traditional lane departure warning systems require visible lane lines.
• Many real-world driving scenarios, especially secondary roads and urban streets, lack clear lane markings.

Purpose of the Study:

• To develop a vision-based method for vehicle localization on roads without lane lines.
• To enable accurate road width and fence-to-fence distance estimation using only RGB imagery.

Main Methods:

• Fusing semantic segmentation and monocular depth estimation to create a semantic 3D point cloud.
• Retaining road points and roadside structures (fences/walls) for spatial analysis.
• Computing road width and fence-to-fence distance along the planned trajectory.

Main Results:

• Successfully demonstrated vehicle localization and road geometry estimation in environments lacking lane lines.
• Quantitative evaluation on Munich street images with road-fence structures.
• Qualitative validation on the Cityscapes dataset (Stuttgart sequence) in a standard urban setting.

Conclusions:

• The proposed method effectively locates vehicles on roads without lane markings.
• The system provides complementary road width and fence-to-fence distance measurements.
• Open-source software is released to benefit the research community.