一种基于残差结构的车道线检测方法

A lane line detection method based on residual structure

提出了一种基于车道线特征的残差因子分解网络实现精确车道线分割的方法,该法采用笔者所提的语义分割网络实现车道线语义分割,通过编码器提取车道线的特征信息,再使用解码器恢复图像信息。在编码器中增加的残差层能更好地处理边缘信息与相似信息,提取到更多的特征信息。用霍夫线拟合方法组成一条可视化的车道线。训练时先对车道线分割训练集进行增强,使用对抗生成网络对公开数据集进行数据增强,自动实现白天到夜晚的转换,生成弱光照场景下的图片,提高训练数据的泛化性。实验证明:笔者算法在保持速度的前提下,能够大大提高分割准确率,与其他车道线分割算法相比,CULane数据集的准确率可提高到74.7%。

A residual factorization network based on lane line features is proposed to achieve precise lane line segmentation.The semantic segmentation network is proposed to achieve lane line semantic segmentation,the feature information of the lane line is extracted through the encoder,and then the decoder is used to restore the image information.The residual layer added in the encoder can better process edge information and similar information,and extract more feature information.The Hough line fitting method is used to compose a visualized lane line.In the training process,the lane line segmentation training set is enhanced firstly,then the generative adversarial network(GAN)is used to enhance the public data set.It can automatically realize the conversion from day to night,generate pictures under weak lighting scenes,and improve the generalization of training data.Experimental analysis proves that the algorithm can greatly improve the segmentation precision while maintaining the speed.Compared with other lane line segmentation algorithms,the segmentation precision on the CULane data set is increased to 74.7%.