Layer-wise feature refinement for accurate three-dimensional lane detection with enhanced bird's eye view transformation

Hao Ren, Mingwei Wang, Yanyang Deng, Wenping Li, Chen Liu

Abstract: Three-dimensional (3D) lane line detection from images is a fundamental yet challenging problem in autonomous driving, with existing methods often lacking scene robustness and computational efficiency. We propose a new method for 3D lane line detection using a simple and efficient view transformation module and layer-wise refined bird’s eye view (BEV) features. Our approach introduces a dual-branch perspective transformation module combining deformable convolution and perspective relationship modules to enhance robustness across diverse scenes. Additionally, we design a cross-attention-based view transformation module with spatial position encoding and BEV spatial query to improve detail learning and transformation effectiveness. Our method further refines BEV features layer by layer to fully exploit multi-level information. Experimental results on two datasets demonstrate the superiority of our approach, showing a significant increase in F1-score compared to existing methods.