基于改进UFLDv2的道路线检测方法

Lane Detection Method Based on Improved UFLDv2

摘要:

在自动驾驶领域中车道线检测算法主要面临检测效率低下，严重遮挡，光照欠佳等问题。基于此应用场景，本文提出了一种基于改进UFLDv2网络的道路线检测方法。本方法在特征提取阶段使用我们设计全新的层结构如SNA与SNB，同时引入轻量化注意力机制CBAM，使用有序数学期望损失函数函数与原始交叉熵损失函数的混合损失函数，有效平衡了FLOPs与检测准确度。最后使用混合锚点检测系统对车道线进行高效的检测，通过在CULane数据集上进行大量实验。结果表明我们改进后的方法所需计算量（FLOPs）降低了57.5%，同时检测精度（总F1参数）提升了0.7%。与其他的主流道路线检测网络相比，改进后的网络在极端情况下的车道线检测取得了很好的效果。

Abstract:

In the field of autonomous driving, lane detection algorithms face challenges such as low detection efficiency, severe occlusion, and poor lighting conditions. To address these issues, this paper proposes a lane detection method based on an improved UFLDv2 network. In the feature extraction stage, we designed new layer structures such as SNA and SNB, introduced the lightweight attention mechanism CBAM, and used a mixed loss function combining an ordered expectation loss and the original cross-entropy loss, effectively balancing FLOPs and detection accuracy. A hybrid anchor detection system was then used for efficient lane detection, and extensive experiments were conducted on the CULane dataset. The results show that our improved method reduces FLOPs by 57.5% and improves detection accuracy (overall F1 score) by 0.7%. Compared to other mainstream lane detection networks, the improved network performs exceptionally well in extreme lane detection scenarios.