Open Access

ARTICLE

Direction-Curvature Aware Feature Integration for Robust Lane Detection

Ahtisham Waheed¹, Yunfie Yin^1,*, Abu Fatema Mohammad Abdun Noor², Md Imam Ahasan¹, Kah Ong Michael Goh^3,*, S. M. Hasan Mahmud^2,*, Umar Rashid⁴

1 College of Computer Science, Chongqing University, Chongqing, China
2 Department of Software Engineering, Daffodil International University, Dhaka, Bangladesh
3 Center for Image and Vision Computing, COE for Artificial Intelligence, Faculty of Information Science & Technology, Multimedia University, Jalan Ayer Keroh Lama, Melaka, Malaysia
4 Department of Computer Science, COMSATS University Islamabad, Islamabad, Pakistan

* Corresponding Authors: Yunfie Yin. Email: ; Kah Ong Michael Goh. Email: ; S. M. Hasan Mahmud. Email:

Computers, Materials & Continua 2026, 89(1), 48 https://doi.org/10.32604/cmc.2026.083456

Received 04 April 2026; Accepted 09 May 2026; Issue published 13 August 2026

Abstract

Robust lane detection is a fundamental perception task for autonomous driving and Advanced Driver Assistance Systems. However, it remains challenging in real-world environments due to degraded lane markings, complex road topologies, occlusions, and adverse illumination conditions. This work aims to improve lane detection robustness by explicitly modeling lane geometric properties while preserving end-to-end efficiency. We propose a direction-curvature aware lane detection framework that integrates a novel Direction-Curvature Aware (DCA) attention module into an anchor-based architecture. The DCA module enables tangent-aligned feature aggregation guided by learned direction fields and curvature-consistent attention. In addition, we introduce a direction-aware optimization objective termed Directional Lane IoU (DLIoU) to enforce directional consistency between predicted and ground-truth lanes during training. Extensive experiments on the CULane and TuSimple benchmarks demonstrate that the proposed method consistently outperforms state-of-the-art approaches, achieving notable improvements on CULane under challenging conditions including occlusion, strong curvature, shadows, and night-time scenes, and consistent gains on TuSimple, while maintaining competitive inference speed. These results confirm that explicitly incorporating direction and curvature information into both feature learning and optimization leads to more accurate and robust lane detection in complex driving environments.

---

Keywords

---