Slope-embedded ViT-based model for lane line detection under occlusions

Deep learning-based lane line detection has garnered substantial success in common scenarios. However, detecting lane lines under conditions of severe occlusion, where visual cues are largely absent, remains a considerable challenge. To address this issue, we propose a cutting-edge strategy that utilizes an enhanced Vision Transformer (ViT) for the de-occlusion of lane lines. Our approach significantly improves the accuracy of lane line detection by integrating a fused feature map with prior knowledge. Specifically, we refine the ViT model by employing overlapping patches technology to reconstruct occluded lane lines from the input image. Subsequently, we extract the feature maps from the model and integrate them with slope and category information pertaining to the lane lines, facilitating more robust and accurate lane line detection. Additionally, we introduce an innovative sensitivity loss function that evaluates not only pixel value errors but also spatial discrepancies between pixels. We assessed our strategy on three benchmark datasets: TuSimple, CULane, and CurveLanes. Our results demonstrate that our approach outperforms existing methods in terms of accuracy and F1-score on all these datasets.