With the rapid development of autopilot technology and various types of sensor, high-precision maps containing a large amount of information for assisting driving have been proposed. The standard lane line detection algorithm relies on the robust estimation of visible lane line markers from a camera image using vision and image processing algorithms. Although the recognition and detection technology for road marking lines is relatively mature, some problems still exist, such as poor detection accuracy and unsatisfactory real-time performance. To solve the problems of the poor robustness and low running speed of the current lane detection methods in complex environments, in this study, we improve current lane detection methods from the perspective of semantic segmentation and propose a DC-VGG-SAD network (VGG: visual geometry group), in which dilated convolution (DC) is used to reduce the complexity of the network to ensure detection accuracy. Furthermore, adding self-attention distillation (SAD) makes the information transmission faster. The proposed network was experimentally evaluated using two large-scale datasets. It was found that when dealing with lane lines in complex environments, the network offers higher detection accuracy and detection speed than most current mainstream networks.