Lane line detection based on cross-convolutional hybrid attention mechanism

• Published in: Scientific reports Vol. 15; no. 1; pp. 16172 - 15
• Main Authors: Wen, Jianping, Zhao, Zhuang, Wang, Chenze, Sun, Ze, Xu, Chao
• Format: Journal Article
• Language: English
• Published: London Nature Publishing Group UK 09.05.2025; Nature Publishing Group; Nature Portfolio
• Subjects: 639/166/988; 639/705/117; Accuracy; Algorithms; Attention mechanism; Autonomous driving; Clustering; Cross-Convolution; Datasets; Design; Humanities and Social Sciences; Lane line detection; Localization; multidisciplinary; Neural networks; Science; Science (multidisciplinary); Attention mechanism; Lane line detection; Autonomous driving; Cross-Convolution; Neural networks
• ISSN: 2045-2322; 2045-2322
• DOI: 10.1038/s41598-025-01167-z

In order to enhance the accuracy and robustness of lane line recognition in dynamic and complex environments, this paper proposes a lane line detection model based on a cross-convolutional hybrid attention mechanism (CCHA-Net). Unlike traditional approaches that separately employ channel and spatial attention, our proposed mechanism integrates these modalities through cross-convolution, thereby enabling cross-group feature interaction and dynamic spatial weight allocation. This novel integration not only improves the continuity of elongated lane features but also enhances the model’s ability to capture long-range dependencies in challenging scenarios. Additionally, this paper designs a lightweight message-passing module that employs dual-branch multi-scale convolutions to achieve cross-spatial domain feature fusion while reducing the number of parameters. Experimental results demonstrate that CCHA-Net achieves an F1 score of 80.2% on the CULane dataset and an accuracy of 96.8% on the TuSimple dataset, effectively enhancing lane line recognition accuracy in ever-changing and intricate environments.