Lane Marker Detection Based on Multihead Self-Attention

Abstract

Lane mark detection is an important task for autonomous driving. Many researchers have proposed many models. But the driving environment is much more complex, especially for some challenging scenarios, such as vehicle occlusion, severe mark degradation, heavy shadow, and so on. It is difficult to detect lane mark in a limited local receptive field under the above scenarios. For that reason, we propose a lane mark detection network based on multihead self-attention. It can find spatial relationships among lane mark points in the global viewpoint and enlarge its feature map’s receptive field equally. For further extracting global and contextual features, it fuses global information and local information together to predict classification and location regression. Finally, it can promote accuracy of lane mark detection greatly especially in challenging scenarios. In the TuSimple benchmark, its accuracy is 95.76% overwhelming all other methods, and its FPS is 170.2, which is the second-highest one. In CULane benchmark its F1 achieves 75.55% and FPS reaches 170.5. Both of them are the highest compared to other methods. Our proposed model establishes a new state-of-the-art among real-time methods.