Topo-Boundary: A Benchmark Dataset on Topological Road-Boundary Detection Using Aerial Images for Autonomous Driving

Zhenhua Xu, Zhenhua Xu 0003, Yuxiang Sun, Ming LiuIEEE RA-L 2021 · DOI

TL;DR

This paper introduces "Topo-boundary," the first large-scale benchmark dataset for topological road-boundary detection using aerial imagery, addressing a critical need for robust autonomous driving systems. They also propose a novel entropy-based metric (ECM) for evaluating topological correctness, demonstrating its superiority over existing methods.

What Problem Does This Paper Address?

Autonomous driving systems rely heavily on accurately identifying road boundaries to ensure vehicles stay within drivable areas and maintain safety. While many existing methods focus on "online" detection using vehicle-mounted sensors (cameras, LiDAR), these approaches often struggle with occlusions and limited onboard computational resources. This paper tackles the problem of "offline" road-boundary detection using aerial images. This approach offers advantages like alleviating occlusion issues and providing data suitable for annotating high-definition (HD) maps. A key challenge in this domain is the lack of a publicly available, comprehensive dataset specifically designed for topological road-boundary detection, hindering research progress. Furthermore, evaluating the topological correctness of detected boundaries, beyond simple pixel-level accuracy, is crucial but lacks standardized, robust metrics.

Methods

The authors introduce the Topo-boundary dataset, which consists of 25,295 aerial image patches (1000x1000 pixels, 4-channel) derived from the NYC Planimetric Database. For each patch, they generate 8 types of ground-truth labels, including binary segmentation, instance maps, and orientation maps, to support various detection approaches.

They also propose a new entropy-based connectivity metric (ECM) for evaluating topological correctness. This metric aims to better handle noise and outliers compared to existing methods like APLS and a naive connectivity metric. ECM assigns higher weights to longer predicted road boundaries that accurately approximate ground truth, thus penalizing disconnections more effectively.

The paper also implements and evaluates several baseline models, including:

  • Segmentation-based baselines: Naive (U-net + filtering), DeepRoadMapper, and OrientationRefine.
  • Graph-based baselines: RoadTracer, VecRoad, ConvBoundary, DagMapper, and iCurb.
  • A newly proposed imitation-learning-based baseline (enhanced-iCurb), which builds upon their previous work (iCurb) by leveraging orientation maps to generate more stable and predictable training labels, improving graph quality.

Datasets

  • NYC Planimetric Database: The source data for creating the Topo-boundary dataset. It contains 2,147 large aerial image tiles with road-boundary ground-truth labels.
  • Topo-boundary Dataset: A newly introduced benchmark dataset comprising 25,295 aerial image patches (1000x1000 pixels, 4-channel) with 8 types of ground-truth labels for topological road-boundary detection. The dataset is split into training, validation, testing, and pretraining sets.

Key Insights and Ideas

  • Aerial imagery is a viable and advantageous source for offline road-boundary detection: It overcomes occlusion issues inherent in vehicle-mounted sensors and is readily available for large-scale mapping.
  • Topological correctness is paramount for road-boundary detection: Simple pixel-level accuracy is insufficient; understanding connectivity and structure is crucial for safe autonomous driving.
  • A dedicated benchmark dataset is essential: The lack of such a dataset was a significant bottleneck, and Topo-boundary aims to fill this gap.
  • Entropy-based metrics offer a more robust evaluation: ECM provides a principled way to assess connectivity, outperforming naive and path-based metrics in handling noise and disconnections.
  • Imitation learning can be effectively adapted for road-boundary detection: The enhanced-iCurb baseline demonstrates how leveraging auxiliary information like orientation maps can lead to more stable and accurate graph generation.

Key Results

  • The Topo-boundary dataset contains 25,295 patches, each with 8 ground-truth labels, making it suitable for diverse research approaches.
  • The proposed ECM metric shows better performance in evaluating topological correctness, especially in scenarios with disconnections, compared to naive and APLS metrics.
  • The enhanced-iCurb baseline achieves superior results, particularly in terms of connectivity and overall graph quality, compared to other baselines, including the original iCurb.
  • Qualitative results (Fig. 6) demonstrate the visual improvements of the proposed method and highlight the challenges faced by other baselines in complex scenarios.
  • Quantitative results (Table I) show that while segmentation-based methods achieve good pixel-level F1-scores, graph-based methods generally perform better on connectivity metrics. The enhanced-iCurb consistently achieves strong performance across various metrics.

Limitations and Future Work

  • Limitations:
    • The dataset was created by filtering out patches with inappropriate annotations, such as those without road boundaries, with intersection points, or with very complex scenarios like overlapping interchanges. This means the current dataset might not fully represent the challenges of real-world applications.
  • The paper acknowledges that the patches removed due to complexity could be considered a limitation for real-world applications.
  • Future Work:
    • The authors plan to assign prediction difficulty scores to further label the dataset, enabling networks to adapt to varying levels of complexity.
  • They intend to consider the patches removed in this paper for future real-world applications, expanding the dataset's scope.

Summary updated

BibTeX
@article{xu2021topo,
  author = {Zhenhua Xu and Zhenhua Xu and Yuxiang Sun and Ming Liu},
  title = {{Topo-Boundary: A Benchmark Dataset on Topological Road-Boundary Detection Using Aerial Images for Autonomous Driving}},
  journal = {IEEE RA-L},
  year = {2021},
  doi = {10.1109/lra.2021.3097512},
}
Figure