Aiming at the problem of poor adaptability of existing lane detection algorithms, a new lane detection algorithm for complex traffic scenes is proposed. Firstly, a lightweight feature pyramid encoding network (FPENet) is used to achieve the semantic segmentation of lane lines to make a good trade-off between accuracy and speed. Secondly, the lightweight Mean shift clustering algorithm is used to cluster the mask lane lines to reduce the calculation cost. Finally, the improved random sample consensus (RANSAC) algorithm is used to implement the lane segmentation. Experimental results show that the algorithm can adapt to various complex road scenes, and the detection accuracy on the self-produced Tusimple extended data set reaches 97.64%, and the detection speed of single frame image reaches 50 frames per second, meeting the detection requirements of accuracy and real-time.