GDALaneNet: A feature fusion strategy balances global awareness and detail accuracy in lane detection

Abstract

Lane detection serves as a core component in autonomous driving technology, forming the foundation for crucial functions such as vehicle autonomous navigation, path planning, and obstacle avoidance. With the continuous advancement of deep neural networks, lane detection algorithm models have seen significant improvements in accuracy, robustness, and real-time performance. However, these models still face challenges posed by the lack of visual cues, such as adverse lighting conditions and occlusion issues. Therefore, to adapt to complex and variable road environments and achieve accurate and efficient lane detection, a new lane detection model named GDALaneNet, which integrates local and global information, has been explored. Through a dual-stream pathway, we combine the ROI features aggregated with global context information with the input features to obtain prior knowledge of lane lines in the image, forming initial proposals and enhancing the model's real-time detection capability. Subsequently, we iteratively refine the proposal features using features from various levels to improve the completeness of the initial proposals, thereby achieving accurate lane detection. Experimental results on three benchmark datasets demonstrate that our method achieves an F1 score of 79.8% on the CULane dataset with a real-time inference speed of over 200 FPS, and an F1 score of 97.93% on the Tusimple dataset, showcasing improvements in both speed and accuracy. On the LLAMAS dataset, F1 score reached 97.1%, the recall rate and accuracy have been effectively improved.