AC-UNet: lane line detection based on U-Net network fusion attention mechanism and cross convolution

Abstract

To address the issue that existing lane line detection algorithms are insufficient in feature extraction under complex road conditions, resulting in poor detection accuracy, based on U-Net, this paper proposes a semantic segmentation network that combines attention mechanism and cross convolution (AC-UNet). The network first uses the special structure of cross-convolution in the encoder to extract the lane line features, allowing the network to keep more edge structure information in different gradient directions. Second, during the AC-UNet generation process, a Feature Refinement Module (FRM) is introduced that selectively emphasizes lane line features by learning global context information, allowing the network to be more sensitive to lane line features. It can improve the transmission of important information while suppressing noise and irrelevant data. The AC-UNet model combines image target information extraction and localization while also addressing the issue of insufficient feature information extraction. Finally, experiments on the TuSimple dataset show that the algorithm in this paper can achieve good detection results even when the lane lines are occluded, with a comprehensive evaluation index F1-measure of up to 90.98%, which improves by 6.53% in comparison to U-Net.