融合ResNeXt与坐标注意力的车道线检测方法

Lane Line Detection Method Integrating ResNeXt and Coordinate Attention

摘要: 随着智能驾驶与ADAS的发展，车道线检测对路径规划和车道保持等模块至关重要，亟需兼顾高精度与实时性的轻量化方法。现有传统方法对遮挡和光照变化鲁棒性不足，深度学习分割和BEV/3D等方法又存在计算开销大或部署成本高的问题，基于行锚的UFLD方法虽具速度优势，但在长距离连续性建模和复杂场景鲁棒性方面仍有提升空间。本文在UFLD框架上引入ResNeXt-50主干网络与坐标注意力机制，并结合分类、结构和分割三种损失共同约束车道线的平滑性与刚性。在CULane数据集上，所提方法取得72.51% F1-score和211.42 FPS，相比其它车道线检测方法达到了精度与速度上平衡。在遮挡、高光、阴影、弯道及夜间等典型场景中，可视化结果显示该方法的漏检与误检数量少于对比方法。实验结果表明，本文所提方法在大幅提升推理速度的同时保持了具有竞争力的检测精度，可用于复杂交通环境下的实时车道线检测。

Abstract: With the development of intelligent driving and ADAS, lane line detection is crucial for modules such as path planning and lane keeping, and there is an urgent need for lightweight methods that balance high precision and real-time performance. The existing traditional methods have insufficient robustness against occlusion and illumination changes. Methods such as deep learning segmentation and BEV/3D have problems of high computational overhead or high deployment cost. Although the UFLD method based on row anchors has a speed advantage, there is still room for improvement in long-distance continuous modeling and robustness in complex scenes. In this paper, the ResNeXt-50 backbone network and coordinate attention mechanism are introduced on the UFLD framework, and the smoothness and rigidity of lane lines are jointly constrained by the three losses of classification, structure and segmentation. On the CULane dataset, the proposed method achieved a score of 72.51% F1-score and 211.42 FPS, achieving a balance between accuracy and speed compared with other lane line detection methods. In typical scenarios such as occlusion, highlights, shadows, curves and at night, the visualization results show that the number of missed detections and false detections of this method is less than that of the comparison method. The experimental results show that the method proposed in this paper not only significantly improves the reasoning speed but also maintains competitive detection accuracy, and can be used for real-time lane line detection in complex traffic environments.