Robust Lane Detection in Autonomous Driving Using Gated Convolutional Networks with Spatial Attention

• Proposed Gated CNN with Spatial Attention enhances lane feature focus. • Hybrid Dice-Focal-Tversky Loss improves thin lane boundary segmentation. • Outperforms six methods on BDD100K, CULane, and TUSimple benchmarks. Lane detection plays an important role in modern self-driving systems, directly bearing on vehicle control, navigation stability, and road safety. All autonomous steering, adaptive cruise control, and lane-keeping assistance were reliant on lane detection. CNN-based models, particularly with encoder-decoder architectures like UNet and SegNet, have inherent limitations in their performance. These limitations become apparent in challenging conditions such as varying lighting, occlusions, faded markings, tight corners, and urban clutter, which can negatively impact their accuracy and robustness. Thus, the present study, introduce a novel deep learning method that utilizes a spatial attention mechanism in combination with Gated Convolutional Neural Networks (Gated CNNs) to mitigate these challenges. Gated convolutional layers enable adaptive feature selection by suppressing irrelevant background information, while the spatial attention module enhances localization of lane-relevant regions. Moreover, a hybrid dice-focal-Tversky loss function is employed to attenuate class imbalance and improve thin lane boundary segmentation. The proposed model was evaluated on three extensive benchmark datasets: BDD100K, CULane, and TUSimple, which collectively represent a wide range of real-world circumstances, including nighttime driving, lane occlusions, and crowded driving conditions. As a result, the model consistently outperformed multiple baseline approaches, in terms of accuracy, F1-score, and intersection over union (IoU) across segmentation experiments. The current framework demonstrates that Gated CNN with Spatial Attention provides a more robust, accurate, and generalizable lane detection approach, applicable in real-time autonomous driving scenarios.