Abstract: Lane line detection, an important prerequisite to ensure the safety of assistant driving and automatic driving, plays a very important role in assistant driving and automatic driving. At present, given the weight-sharing convolutional neural network (CNN) and the reduced training parameters, CNN can automatically learn and extract features, and can be widely used in the image segmentation and recognition and other fields. Based on the characteristics of lane detection, this paper improves the symmetric convolution kernel to the asymmetric kernel CNN (AK-CNN) structure, which further reduces the computation of CNN network and enhances the speed of lane detection.