Numerous groups have conducted many studies on traffic lane detection. However, most methods detect lane regions by color feature or shape models designed by human. In this paper, a traffic lane detection method using fully convolutional neural network is proposed. To extract the suitable lane feature, a small neural network is built to implement feature extraction from large amount of images. The parameters of lane classification network model are utilized to initialize layers' parameters in lane detection network. In particular, a detection loss function is proposed to train the fully convolutional lane detection network whose output is pixel-wise detection of lane categories and location. The designed detection loss function consists of lane classification loss and regression loss. With detected lane pixels, lane marking can be easily realized by random sample consensus rather than complex post-processing. Experimental results show that the classification accuracy of the classification network model for each category is larger than 97.5%. And detection accuracy of the model trained by proposed detection loss function can reach 82.24% in 29 different road scenes.