AI Automated Driving Assistance system using Lane detection with RNN

Engineering notes

Evaluations on public benchmarks show improved accuracy and robustness across diverse environmental conditions.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

This Lane detection is vital for autonomous driving, impacting vehicle guidance and safety. Traditional 2D methods face projection errors, while 3D lane detection advances focus on geometric curve propagation or pixel-level instance segmentation. We propose a hybrid framework combining a transformer-based curve propagation module with an affinity field-based clustering module. The transformer module uses dynamic curve queries to predict 3D lane parameters without explicit view transformation, while the affinity module employs horizontal and vertical affinity fields to cluster lane pixels into distinct instances robustly. Evaluations on public benchmarks show improved accuracy and robustness across diverse environmental conditions. We detail the network architecture, loss formulation, and ablation studies, and explore future extensions. This approach enhances 3D lane detection for safer autonomous driving.

Links and sources

• Official / arXiv page

Comments