MultiLane-CtransFuseNet: A Structure-Aware Lane Detection Framework with Curvature-Constrained Transformer

Engineering notes

The proposed architecture achieves high detection accu-racy while maintaining real-time performance. Extensive experiments on TuSimple and CULane benchmarks demonstrate that our method achieves 96.63% and 77.32% accuracy, respectively, and runs at 134 FPS, confirming its robustness and deployment potential in real-world scenarios.

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

Lane detection plays a critical role in autonomous driving and Advanced Driver Assistance Systems (ADAS). Traditional approaches typically treat it as a pixel-wise segmentation task, relying on low-level features such as color, gradient, and texture, which struggle under complex conditions like occlusions, varying illumination, and multi-lane scenarios [8]. To address these challenges, we propose MultiLane-CtransFuseNet, a structure-aware framework that integrates lightweight Convolutional Neural Networks (CNNs) with Transformers. Inspired by human visual perception, the network employs spatial–channel dual attention and global context modeling to enhance feature specificity and structural consistency. By incorporating the query-based mechanism from DETR, we reformulate lane detection as a structured object modeling task, enabling better topological understanding of lane layouts [1]. We further design a novel structure-aware loss that jointly enforces local continuity and global curvature regularization, alongside an attention alignment mechanism to guide the Transformer’s focus toward lane-relevant regions. The proposed architecture achieves high detection accu-racy while maintaining real-time performance. Extensive experiments on TuSimple and CULane benchmarks demonstrate that our method achieves 96.63% and 77.32% accuracy, respectively, and runs at 134 FPS, confirming its robustness and deployment potential in real-world scenarios.

Links and sources

• Official / arXiv page

Comments