Monocular BEV Perception of Road Scenes via Front-to-Top View Projection

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• High-definition (HD) map reconstruction is vital for autonomous driving systems.
• Current LiDAR-based methods are costly and computationally intensive.
• Existing camera-based approaches often suffer from distortion and data loss due to separate road segmentation and view transformation.

Purpose of the Study:

• To develop an efficient framework for reconstructing local bird's-eye view (BEV) maps from monocular front-view images.
• To improve road layout and vehicle occupancy estimation for autonomous navigation.
• To overcome limitations of existing camera-based HD map reconstruction techniques.

Main Methods:

• A novel front-to-top view projection (FTVP) module is proposed, incorporating cycle consistency to enhance view transformation and scene understanding.
• Multi-scale FTVP modules are utilized to propagate low-level feature information, reducing spatial deviation in object localization.
• The framework processes monocular front-view images to generate BEV maps.

Main Results:

• The proposed method achieves performance comparable to state-of-the-art methods in road layout estimation, vehicle occupancy estimation, and multi-class semantic estimation.
• The framework demonstrates superior computational efficiency compared to existing approaches.
• Experiments on public benchmarks validate the effectiveness of the novel approach.

Conclusions:

• The developed framework offers an efficient and effective solution for HD map reconstruction using monocular camera input.
• The FTVP module significantly improves view transformation and scene understanding in autonomous driving contexts.
• This method presents a promising alternative to LiDAR-based mapping, enhancing accessibility and reducing computational load.