The lane detection is a key problem to solve the division of derivable areas in unmanned driving, and the detection accuracy of lane lines plays an important role in the decision-making of vehicle driving. Scenes faced by vehicles in daily driving are relatively complex. Bright light, insufficient light, and crowded vehicles will bring varying degrees of difficulty to lane detection. So we combine the advantages of spatial convolution in spatial information processing and the efficiency of ERFNet in semantic segmentation, propose an end-to-end network to lane detection in a variety of complex scenes. And we design the information exchange block by combining spatial convolution and dilated convolution, which plays a great role in understanding detailed information. Finally, our network was tested on the CULane database and its F1-measure with IOU threshold of 0.5 can reach 71.9%.