Introduction
High-definition (HD) maps serve as the "digital rails" for autonomous vehicles (AVs), providing centimeter-level precision for lane boundaries, stop lines, and traffic symbols. While ground-level sensors like LiDAR are standard, they are labor-intensive to scale and limited by field-of-view constraints. Aerial imagery offers a scalable alternative, yet extracting thin, precise lane markings remains challenging due to obstructions (trees, shadows), scale variance, and data scarcity. This study evaluates 12 state-of-the-art deep learning models to automate this process, highlighting the transition from traditional Convolutional Neural Networks (CNNs) to Transformer-based architectures.
Methodology
The research compares 12 semantic segmentation models categorized into CNN-based architectures (e.g., FCN, DeepLabV3+, SegNeXt) and Transformer-based architectures (e.g., SegFormer, Swin). The models were evaluated based on their computational complexity (FLOPs) and parameter counts to determine their suitability for real-world deployment.
Training Strategy
A three-stage transfer learning pipeline was employed:
- Pre-training: Initializing weights using ImageNet.
- Domain Transfer: Training on the SkyScapes dataset (13cm resolution imagery of Munich) to learn general aerial road features.
- Fine-tuning: Refinement on an enhanced Waterloo Urban Scene dataset (12cm resolution imagery of Ontario), manually labeled with 14 classes including crosswalks, turn signs, and traffic islands.
Results and Analysis
The study reveals a significant performance gap between architectures. Transformer-based models, particularly SegFormer, outperformed traditional CNNs by leveraging self-attention mechanisms to capture long-range spatial relationships. This "global view" allows Transformers to reconstruct lane markings even when partially obscured by shadows or vegetation—a task where local-feature-focused CNNs often struggle.
Key Performance Metrics
- Top Performer: SegFormer was the clear winner, achieving a mean Intersection over Union (mIoU) of 76.11% and an F1 score of 85.35% on the Waterloo dataset.
- Efficiency vs. Accuracy: While Transformers led in accuracy, MobileNetV3 demonstrated high efficiency with significantly lower FLOPs and parameters, making it a candidate for resource-constrained real-time applications despite lower mIoU.
- Recall vs. Precision: Most models exhibited higher Recall than Precision, indicating a tendency to "over-predict" lane markings by misclassifying road edges or sidewalk boundaries as lanes.
Conclusion
The shift toward Transformer-based architectures represents a major advancement in aerial lane mapping. By effectively handling occlusions and maintaining global context, these models provide the reliability needed for HD map generation. Future work will explore the use of Generative AI (Diffusion Models) to synthesize training data for rare environmental conditions and the development of loss functions tailored to address class imbalance in thin feature segmentation.
Summary updated