Anchor3DLane++: 3D Lane Detection Via Sample-Adaptive Sparse 3D Anchor Regression.

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Machine Learning

Background:

• Traditional 3D lane detection methods rely on Inverse Perspective Mapping (IPM) to Bird-Eye-View (BEV), which suffers from flat ground assumptions and information loss.
• Existing BEV-free methods lack structured modeling for 3D lanes, hindering their performance compared to BEV-based approaches.

Purpose of the Study:

• To propose a novel and effective BEV-free monocular 3D lane detection method.
• To improve the accuracy and robustness of 3D lane detection by introducing structural representations and adaptive anchor generation.

Main Methods:

• Anchor3DLane++ utilizes 3D lane anchors as structural representations for direct prediction from Front-View (FV) features.
• A Prototype-based Adaptive Anchor Generation (PAAG) module dynamically creates sparse 3D anchors.
• An Equal-Width (EW) loss function is employed for lane regularization, leveraging their parallel nature.
• Camera-LiDAR fusion is explored to enhance performance by combining complementary sensor data.

Main Results:

• Anchor3DLane++ demonstrates superior performance compared to existing state-of-the-art methods on three benchmark datasets.
• The proposed method effectively addresses limitations of IPM and improves 3D information estimation.
• The integration of PAAG and EW loss contributes to more accurate and robust 3D lane detection.

Conclusions:

• Anchor3DLane++ offers a promising BEV-free approach for monocular 3D lane detection.
• The method's structural representation and adaptive generation of anchors significantly advance the field.
• Future work can explore further enhancements through sensor fusion and refined loss functions.