IFPNet: Integrated Feature Pyramid Network with Fusion Factor for Lane Detection

Abstract

Lane detection is a basic but challenging task in autonomous driving systems. With a combination of high-level and low-level information, early studies of lane detection have achieved promising results in some scenes. However, achieving better performance is still an urgent need for complex and diverse road conditions. We assume that learning and balancing the finer-scale features and global semantics is one of the keys to improving lane detection performance under these road conditions. In this paper, we propose an integrated feature pyramid network with fusion factor (IFPNet) for better hierarchical information learning and balancing, where a novel FPN structure named Integrated Feature Pyramid (IFP) is proposed for better hierarchical information integration. Classification Fusion Factor (CFF) is also utilized for the balance of hierarchical information. Moreover, we design the regression IoU (RIoU) loss for curve regression, which measures the overlap of the predicted and ground truth lane lines more effectively. We conduct experiments on three benchmark datasets of lane detection and achieve state-of-the-art results with high accuracy and efficiency.