基于DSA-UFLD模型的车道线检测算法

Lane line detection algorithm based on DSA-UFLD model

车道线检测是保证自动驾驶安全性与稳定性的关键,为提高车道线检测的准确性,本文基于UFLD(Ultra Fast Structure-aware Deep Lane Detection)算法,结合DenseNet-121网络和空间注意力(Spatial Attention)机制,设计了一种DSA-UFLD模型实现车道线检测。在图像增强方面,使用图像亮度自适应增强算法提高欠曝图像的清晰度;在网络优化方面,用迁移学习模型DenseNet-121代替ResNet18提取图像特征,利用密集连接加强特征重用,并引入空间注意力机制提取图像的关键信息,其次在上采样中用转置卷积代替双线性插值,通过学习参数,更好地实现解码;在损失函数方面,通过改进结构损失,将车道线约束为二次曲线,改善了弯道场景下车道线的检测效果。实验结果表明,DSA-UFLD算法在保证检测速度的同时,提高了车道线的识别准确率,具有一定的应用价值。

Lane detection is the key to ensure the safety and stability of automatic driving.In order to improve the accuracy of lane detection,this paper designs a DSA-UFLD model to achieve lane line detection based on UFLD(Ultra Fast Structure-aware Deep Lane Detection)algorithm,combined with DenseNet-121 network and Spatial Attention mechanism.In the aspect of image enhancement,the image brightness adaptive enhancement algorithm is used to improve the definition of underexposed images.In the aspect of network optimization,transfer learning model DenseNet-121 is used instead of ResNet18 to extract image features,dense connection is used to strengthen feature reuse,and spatial attention mechanism is introduced to extract key information of the image.Then,transposed convolution is used instead of bilinear interpolation in upsampling,and decoding is better realized by learning parameters.In the aspect of loss function,by improving the structural loss,the lane line is constrained to a quadratic curve,which improves the detection effect of the lane line in the curve scene.The experimental results show that DSA-UFLD algorithm not only ensures the detection speed,but also improves the recognition accuracy of lane lines,which has certain application value.