基于边缘特征融合和跨连接的车道线语义分割神经网络

Lane Semantic Segmentation Neural Network Based on Edge Feature Merging and Skip Connections

无人驾驶中的车道线检测任务需要同时确定车道线的位置、颜色和线型,而现有方法通常仅识别车道线的位置,不识别车道线的类型.为了端到端地解决这一问题,设计了一种语义分割神经网络,将一幅图像中不同车道线分割为不同区域,用每个区域的类别标签表示其对应的车道线类型.首先,在主流的编码器-解码器框架下,构建了一个结构较为简单的基础网络.考虑到边缘特征是车道线检测中的重点,为基础网络的编码器并联了一个边缘特征提取子网络,通过逐层融合边缘特征图和原始特征图增强车道线的特征.边缘特征提取子网络的结构与基础网络的编码器相同,其输入是对车道线图像进行Sobel滤波的结果.此外,编码器和解码器对称位置的卷积层输出的特征图尺寸相同,但具有不同的语义层级.为了更好地利用这一特性,建立从编码器到解码器对称位置的跨连接,在解码器逐层上采样的过程中融合编码器对应尺寸的特征图.在TSD-Lane车道线检测数据集上的实验表明,相比于基础网络,基于边缘特征融合和跨连接的神经网络的分割性能得到了较为显著的提高.该网络具有较好的车道线分割性能,能够在确定车道线位置的同时,区分黄线或白线、虚线或实线.在计算资源充足的前提下,该网络能够达到实时的检测速度.

In autonomous driving,the lane detection task is required to detect the color,the type and the position of each lane.However,most existing methods usually detect the lane positions only,without recognizing the color and the type of each lane.To find an end-to-end solution to this problem,a semantic segmentation neural network is designed.In an image,different lanes are segmented into different regions.The label of each region represents the type of the corresponding lane.First,a rather simple base network is constructed basing on the main-stream encoderdecoder framework.Considered that edge features are important in lane detection,an edge feature extracting subnetwork is parallel connected to the encoder of the base network,enhancing lane features by merging original feature maps with edge feature maps layer by layer.The results of applying the Sobel filter to lane images are fed into the edge feature extracting subnetwork,which shares an identical architecture to the original encoder of the base network.Besides,the feature maps from the symmetrical convolutional layers of the encoder and the decoder have the same size,but their semantic levels are different.In order to make better use of this property,skip connections from the encoder to the decoder are implemented symmetrically,merging the corresponding encoder feature maps to the decoder feature maps in the procedure of upsampling.Experiments on TSD-Lane lane detection dataset demonstrate that the performance of the neural network based on edge feature merging and skip connections is improved rather significantly,compared with the base network.The proposed network provides good performance on lane segmentation,and it is able to detect the color,the type and the position of each lane simultaneously.Under the condition of having enough computational resources,the proposed network can achieve real-time detection.