基于物理几何一致性的车道检测

Lane Detection Based on Physical and Geometric Consistency

摘要: 为应对复杂路况下车道检测的结构鲁棒性挑战，本文提出一种融合物理几何一致性的新方法。现有技术在光照不均、遮挡及标线磨损等场景下，常因缺乏对车道固有几何约束的考量，导致检测结果出现不连续、弯折或跳变等结构性缺陷。为解决此问题，我们设计了多个可微的物理几何先验，包括平滑性、透视一致性、仿射稳定性与长度连续性，并将其作为正则项融入深度学习框架。通过这种方式，模型在优化过程中能够学习并遵守车道的内在几何规范。在多个公开数据集上的实验验证，本方法显著提升了车道检测的鲁棒性，尤其在弯道及不完整标线等挑战性场景中，其性能超越了现有先进方法。

Abstract: To address the challenge of structural robustness in lane detection under complex road conditions, this paper proposes a novel method that integrates physical and geometric consistency constraints. Existing techniques often produce structural defects such as discontinuities, sharp bends, or abrupt changes in detection results when facing uneven lighting, occlusions, and worn markings, primarily due to the lack of consideration for the inherent geometric constraints of lanes. To solve this problem, we design multiple differentiable physical-geometric priors, including smoothness, perspective consistency, affine stability, and length continuity, and incorporate them as regularization terms into a deep learning framework. Through this approach, the model can learn and adhere to the intrinsic geometric principles of lane structures during optimization. Experimental validation on multiple public datasets demonstrates that our method significantly improves the robustness of lane detection, particularly in challenging scenarios such as curves and incomplete lane markings, outperforming existing state-of-the-art methods.