Flexible 3D Lane Detection by Hierarchical Shape Matching

Zhihao Guan, Ruixin Liu, Zejian Yuan, Ao Liu, Kun Tang, Tong Zhou, Erlong Li, Chao Zheng, Shuqi MeiAAAI 2023 · DOI

High-Definition (HD) map construction demands centimeter-level precision in 3D lane detection, a requirement that remains a significant challenge for current autonomous driving systems. Traditional methods often struggle with complex road topologies, local geometric inaccuracies, or the computational overhead of post-processing. To address these issues, the Flexible Hierarchical Lane Detector (FHLD) was introduced at AAAI-23. FHLD is an end-to-end framework that achieves high-precision 3D lane detection by combining global parametric modeling with local segment refinement.

The Challenges of High-Precision Lane Detection

Most existing lane detection benchmarks use permissive metrics that do not meet the 10–30 cm accuracy threshold required for HD maps. Current methodologies typically fall into three categories, each with inherent limitations:

  1. Segmentation-based methods: These require complex clustering post-processing to group pixels into lane instances.
  2. Global parametric methods: While effective at capturing the overall lane structure, they often overlook local geometric errors.
  3. Fixed-grid methods: These struggle with non-standard topologies, such as lane merges, splits, or lanes that do not align with a standard coordinate axis.

The FHLD Architecture

FHLD addresses these limitations through a hierarchical approach that operates across two distinct branches: a global parametric curve branch and a local segment shape branch. This dual-stage process allows the model to capture both the high-level semantics of the lane and the fine-grained local details.

The FHLD architecture overview, showing the flow from BEV maps through the Backbone to the hierarchical Global and Local branches.

1. Global Parametric Curve Branch

To overcome the rigidity of fixed-axis predictions, FHLD utilizes an adaptive reference axis. Instead of predicting lanes relative to a fixed Y-axis, the model identifies the start ($p_s$) and end ($p_e$) points of a lane instance. The line connecting these two points serves as a custom reference axis. The lane is then modeled as a 3D cubic polynomial relative to this axis, implemented via a DETR-like Transformer structure. This flexibility allows the network to handle curved or diagonal lanes without fighting the underlying geometry of the road.

2. Local Segment Shape Branch

To correct local errors inherent in global fitting, the model generates Dynamic Anchor Cells. Points are sampled along the predicted global curve, and local feature maps are cropped around them. Within each cell, the lane is modeled as a "thin bar" defined by a center, length, and heading.

Technical details of the hierarchical approach: (a) the adaptive reference axis defined by start and end points, and (b) the local segment modeling using Gaussian distribution matching.

Rather than standard coordinate regression, FHLD employs Gaussian Matching. The local prediction is optimized by matching a 2D Gaussian distribution (representing the lane segment) to the ground truth using KL-divergence. This approach provides a smoother optimization landscape and makes the model more resilient to noise in the input point cloud.

3. Hierarchical Loss Function

The framework is trained using a composite loss strategy. A Global Shape Matching Loss uses Hungarian matching to pair instance predictions with ground truth, while a Local Shape Matching Loss incorporates curve smoothing to ensure that the sequence of local segments forms a cohesive, continuous line.

Experimental Results and Insights

The researchers evaluated FHLD on two Bird’s-Eye View (BEV) point-cloud datasets: RoadBEV (a new 25k-image dataset featuring complex splits and merges) and sub-KCUD (a subset of the KAIST Complex Urban Dataset). BEV representations were chosen for their immunity to lighting conditions and preservation of 3D geometry.

Key results include:

  • High Precision: On the RoadBEV dataset, FHLD achieved an F1 score of 77.57% at a 10cm threshold, outperforming CondLaneNet by 4% and parametric methods like LSTR by over 5.6%.
  • Geometric Robustness: In scenarios with large curvatures, the adaptive representation maintained a significant lead over traditional fixed-axis models.
  • Real-Time Performance: Despite its hierarchical complexity, FHLD runs at 97 FPS on an NVIDIA RTX 3090, making it highly suitable for real-time deployment.

Conclusion

FHLD represents a shift toward more flexible, hierarchical representations in 3D lane detection. By allowing lanes to define their own coordinate systems and refining global shapes with local Gaussian-matched segments, the model achieves the centimeter-level accuracy necessary for the next generation of HD maps. Future work may involve refining anchor cell hyperparameters and extending the framework to multi-modal sensor fusion inputs.

Summary updated

BibTeX
@inproceedings{guan2023flexible,
  author = {Zhihao Guan and Ruixin Liu and Zejian Yuan and Ao Liu and Kun Tang and Tong Zhou and Erlong Li and Chao Zheng and Shuqi Mei},
  title = {{Flexible 3D Lane Detection by Hierarchical Shape Matching}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i1.25146},
}
Figure