Abstract
Lane detection is a fundamental task for autonomous driving, yet balancing efficiency with holistic geometric understanding remains challenging. Traditional segmentation and point-based methods often suffer from high computational costs or reliance on heuristic anchors. This paper introduces BezierLaneNet, an end-to-end framework that models lanes as parametric cubic Bézier curves. By predicting control points in image space and incorporating a Feature Flip Fusion module to exploit road symmetry, the proposed method achieves state-of-the-art performance on the LLAMAS benchmark. BezierLaneNet maintains a compact footprint of under 10M parameters and operates at over 150 FPS, demonstrating superior robustness in challenging lighting conditions.
Introduction
Lane detection serves as a cornerstone for autonomous navigation, providing critical spatial constraints for vehicle positioning. However, current deep learning approaches face a persistent trade-off between real-time efficiency and a "holistic" understanding of road geometry. Most existing methodologies fall into two categories:
- Segmentation-based methods: These treat lanes as pixel-wise masks. While accurate, they are computationally expensive and require significant post-processing to group pixels into coherent line instances.
- Point-based methods: These detect lanes as discrete points, often relying on thousands of heuristic "anchors" and requiring Non-Maximum Suppression (NMS) to resolve overlapping predictions.
While curve-based methods using polynomials offer a more structured representation, they have historically underperformed. The mathematical instability of abstract polynomial coefficients (e.g., $ax^3 + bx^2 + cx + d$) makes them difficult for neural networks to optimize directly from visual features.
Methodology
BezierLaneNet reimagines lane detection through three primary innovations designed to improve both learnability and structural coherence.
Cubic Bézier Curve Modeling
Rather than regressing abstract coefficients, the network predicts four "control points" for each lane. Because these points reside within the image coordinate space, they provide a more intuitive optimization target for convolutional neural networks (CNNs). The resulting curve is generated by interpolating between these points, creating a "convex hull" that is naturally robust to localized noise and occlusions.
Feature Flip Fusion
Driving environments are inherently symmetric, with left-side lane markings often mirroring those on the right. BezierLaneNet exploits this prior through a Feature Flip Fusion module. The network horizontally flips internal feature maps and fuses them with the original maps using deformable convolutions. This mechanism allows the model to infer the position of faded or occluded lanes by leveraging information from well-marked lanes on the opposite side of the road.
End-to-End Bipartite Matching
Inspired by modern object detection architectures like DETR, the model utilizes a sparse set of lane proposals. These proposals are matched to ground truth labels using the Hungarian algorithm, resulting in a fully differentiable, NMS-free pipeline that maps directly from input images to parametric curves. To ensure the predicted curves align precisely with ground truth, the authors employ a Sampling Loss that calculates error across 100 points along the curve rather than just at the control points.
Experimental Results
The model was rigorously evaluated on three major benchmarks: TuSimple (highway), CULane (urban/challenging), and LLAMAS (large-scale/HD map-based).
Performance and Efficiency
Using a ResNet-34 backbone, BezierLaneNet achieves a processing speed exceeding 150 FPS on a single RTX 2080 Ti. With fewer than 10 million parameters, the architecture is highly suitable for edge deployment.
Accuracy and Robustness
The framework established a new state-of-the-art (SOTA) on the LLAMAS benchmark. On the challenging CULane dataset, it outperformed previous curve-based methods by a significant 6.85% F1 score. Notably, the model demonstrated superior robustness in "Dazzle Light" and "Shadow" categories, where global geometric priors compensate for degraded local pixel textures.
Conclusion and Future Work
BezierLaneNet demonstrates that curve modeling, when properly formulated via Bézier control points and symmetry priors, can outperform traditional segmentation and point-based methods. While highly effective, the authors note that the model currently faces challenges with extreme curvatures (e.g., sharp mountain hairpins) due to dataset bias. Additionally, the Feature Flip Fusion module assumes a standard front-facing camera perspective. Future research will focus on extending these geometric priors to handle a broader range of camera configurations and more complex road topologies.
Summary updated