Lane line detection is a crucial part of autonomous driving technology. To address the problem of high false detection rate, as well as the difficulty in balancing frame rate and accuracy in the autonomous driving scenarios, an end-to-end GNN-Transformer detection framework was developed, in which graph neural network (GNN) was used to enhance the local geometric consistency of lane lines and a Transformer encoder-decoder was employed to complete the global dependency modeling and lane line prediction. In addition, learnable positional encoding was adopted and an optimized curve fitting strategy was introduced to improve the model's adaptability to complex scenarios. The proposed lane line detection method was experimentally verified on the TuSimple dataset, CULane dataset and CARLA simulator. Experimental results on the TuSimple dataset showed that the proposed method achieved a false detection rate of 0.019 2, which was reduced by up to 89% compared with other six methods, including ORANet. Meanwhile, the frame rate remained at 110 frames per second, indicating that the method achieved high detection accuracy, stability and real-time performance. Furthermore, the model was deployed on an RTRC4pro intelligent vehicle, thereby further evaluating the engineering application potential of the proposed method. The research results can provide strong support for the online perception of lane lines and its engineering applications in real-vehicle scenarios.