Compact HD Map Construction via Douglas-Peucker Point Transformer

Ruixin Liu, Zejian YuanAAAI 2024 · DOI

TL;DR

This paper introduces DPFormer, a novel end-to-end framework for constructing compact High-Definition (HD) maps. It utilizes a Douglas-Peucker (DP) point representation to efficiently capture map element geometry with fewer points, combined with a transformer architecture for robust detection and novel constraints for improved accuracy.

What Problem Does This Paper Address?

Autonomous driving systems rely heavily on High-Definition (HD) maps, which provide centimeter-level accuracy and rich semantic information about the road environment. Traditional methods for creating these maps often result in redundant point representations or overly complex curve models, making them inefficient and difficult to maintain. Existing online HD map construction approaches either treat it as a semantic segmentation problem (losing geometric detail) or focus on sub-tasks like lane detection, which can be limited by specific views or shape priors. While recent methods aim for vectorized HD maps, they still face challenges with compact representations (e.g., equidistant points that miss crucial details or Bézier curves that neglect global structure) and ambiguities in point regression due to viewpoint transformations. This paper addresses the need for a more compact, accurate, and robust representation and detection method for HD map elements.

Methods

The core innovation of DPFormer lies in its compact Douglas-Peucker (DP) point representation. Instead of using dense equidistant points or complex Bézier curves, DPFormer approximates map elements with a minimal set of points identified by the Douglas-Peucker algorithm. These DP points are strategically placed to preserve essential geometric structures while discarding redundancy.

The framework is built upon a transformer-based architecture that takes multi-view images as input. A CNN backbone extracts features, which are then transformed into a Bird's Eye View (BEV) representation. This BEV feature map is processed by a cascaded transformer decoder. The decoder uses hierarchical queries to detect map elements.

To enhance robustness and accuracy, DPFormer introduces several novel constraints:

  • Class-agnostic representation: The DP point representation is designed to work for any map element type without prior class knowledge.
  • Position constraint with uncertainty estimation: To mitigate ambiguities arising from viewpoint transformations and potential calibration errors, the model predicts not only point positions but also their uncertainty. This helps in learning more reliable position estimations.
  • Pairwise-point shape matching constraints: These constraints are applied at both single-point and pairwise levels to balance local structural information across different scales and improve the overall shape representation. This is achieved using a Kullback-Leibler divergence loss between predicted and ground-truth shape distributions.
  • Hierarchical matching: A unified approach to instance-level and point-level matching is employed, leveraging the Hungarian algorithm for optimal assignment between predicted and ground-truth elements.

Datasets

The paper uses the nuScenes dataset for both training and evaluation. nuScenes is a large-scale multimodal dataset that includes synchronized camera images and LiDAR sweeps. The authors follow common practice by training on the training set and evaluating on the validation set. They specifically focus on lane dividers, pedestrian crossings, and road boundaries for evaluation.

Key Insights and Ideas

  • Douglas-Peucker for HD Maps: The Douglas-Peucker algorithm, traditionally used for simplifying polylines, proves to be an effective and surprisingly robust method for creating compact and accurate representations of complex HD map elements. This challenges the assumption that more points always equate to better detail.
  • Uncertainty as a Regularizer: Incorporating uncertainty estimation in point regression is crucial for handling the inherent ambiguities in multi-view perception and viewpoint transformations, leading to more reliable predictions.
  • Hierarchical Constraints for Robustness: The combination of instance-level classification, point-level classification, uncertainty-aware position regression, and pairwise shape matching creates a powerful hierarchical supervision system that captures both global structure and local details effectively.
  • Compactness without Sacrificing Accuracy: The DP representation, by design, leads to significantly fewer points per map element, achieving substantial compression ratios without a significant drop in performance, and in many cases, even improving it.

Key Results

DPFormer achieves state-of-the-art (SOTA) performance on the nuScenes dataset.

  • Performance: The method outperforms previous SOTA methods, including MapTR, with a higher mAP. Specifically, it achieves a mAP of 60.8% with a ResNet50 backbone trained for 150 epochs.
  • Compactness: DPFormer achieves impressive compression ratios, using significantly fewer points than other methods. For instance, it achieves a compression ratio of 45.5% in one of the visualized samples, meaning it uses less than half the points of the original annotations. This is a substantial improvement over methods like MapTR, which uses more points for similar or slightly lower performance.
  • Ablation Studies: Extensive ablation studies validate the effectiveness of each component. The DP point representation itself shows significant gains over equidistant points. The uncertainty estimation and pairwise shape matching constraints also contribute positively to the overall performance.

Limitations and Future Work

  • Limitations:
    • The paper notes that for complex shapes, especially boundaries with intricate topological structures, the accuracy of ground-truth DP point generation might be compromised, potentially diminishing the advantages of the DP representation in such specific cases.
  • While DPFormer excels at capturing structures, the paper acknowledges that the "sacrifice in accuracy during ground-truth DP point generation diminishes the advantages of DP points" for certain complex boundary types.
  • Future Work:
    • The authors suggest combining DPFormer with other auxiliary tasks to enhance performance.
  • They also propose extending the framework to downstream tasks such as multi-modality predictions and motion planning, indicating the potential of their compact HD map representation for broader applications in autonomous driving.

Summary updated

BibTeX
@inproceedings{liu2024compact,
  author = {Ruixin Liu and Zejian Yuan},
  title = {{Compact HD Map Construction via Douglas-Peucker Point Transformer}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i4.28160},
}

Loading related content...

Figure