Abstract
Monocular 3D lane detection is a critical component for autonomous driving, yet it faces significant challenges due to perspective distortion and geometric misalignments on non-flat terrain. We present PVALane, an end-to-end framework that addresses these hurdles by replacing traditional dense 3D anchors with sparse, high-quality "Prior Anchors" derived from 2D image predictions. PVALane employs a dual-branch architecture—a Context-aware Front-View (FV) encoder and a Geometric-aware Bird’s-Eye-View (BEV) encoder—and aligns their features using a novel View-Agnostic Feature Alignment (PVFA) module. Evaluated on industry-standard benchmarks OpenLane and ONCE-3DLanes, PVALane achieves state-of-the-art accuracy (63.4% F1 on OpenLane) and impressive real-time performance of up to 108 FPS.
Introduction
Accurate 3D lane detection from a single camera is essential for vehicle localization and path planning. However, existing methods struggle with two primary issues:
- Dense Anchor Inefficiency: Standard models often rely on thousands of pre-defined 3D anchors. This density leads to feature overlap and a high rate of false positives, as many anchors are indistinguishable in the 2D projection.
- Geometric Misalignment: Methods using Inverse Perspective Mapping (IPM) to create a BEV representation assume a flat road. On hills or slopes, this assumption fails, causing misaligned lane detections. Furthermore, the BEV transformation often discards valuable semantic context present in the original front-view image.
PVALane bridges these gaps by combining the semantic richness of the front-view with the geometric consistency of the bird’s-eye-view, utilizing a sparse search space for efficiency.
Methods
The PVALane architecture consists of three core components designed to optimize both accuracy and speed.
Prior Anchor Network (PAN)
Instead of evaluating a dense set of anchors, PAN first predicts lane existence in the 2D image plane. By projecting 3D anchors into 2D and sampling features to calculate "objectness" scores, the model filters out low-probability candidates. Only a sparse set of "Prior Anchors" is retained, significantly reducing the search space and eliminating redundant "ghost" detections.
View-Specific Encoders
The model processes input through two parallel branches:
- FV Context-aware Encoder: Utilizes a Transformer to capture global semantic information, such as roadside landmarks or highway exits, which are often lost in top-down views.
- BEV Geometric-aware Encoder: Employs a pyramidal structure to extract inherent geometric properties, such as lane parallelism, which are more apparent in the BEV space.
Prior-Guided View-Agnostic Feature Alignment (PVFA)
The PVFA module acts as the bridge between views. Using the sparse prior anchors, it samples features from both the FV and BEV encoders. Since these features are tied to the same 3D anchor, they are naturally aligned regardless of terrain irregularities. The fusion of these complementary features allows the model to leverage both appearance and geometry for final 3D lane regression.
Key Insights
- Sparsity for Precision: Pruning the 3D search space using 2D priors avoids the confusion inherent in dense designs, leading to higher precision and fewer false positives.
- Complementary View Fusion: FV features excel at identifying what an object is (semantics), while BEV features are superior for determining where it is (geometry). Simultaneous sampling from both views, guided by 3D anchors, effectively mitigates alignment issues caused by non-flat ground.
Results
PVALane was evaluated on the OpenLane and ONCE-3DLanes benchmarks.
- State-of-the-Art Performance: On OpenLane, PVALane achieved an F1 score of 63.4% with a Swin-B backbone, outperforming established models like LATR and BEV-LaneDet.
- Real-Time Efficiency: With a ResNet-18 backbone, the model reaches 108 FPS, making it highly suitable for deployment on standard autonomous vehicle hardware.
- Robustness: The model showed significant gains in challenging scenarios such as "Extreme Weather" and "Night," where traditional single-view methods often struggle.
Conclusion
PVALane demonstrates that sparse, prior-guided anchors and multi-view feature alignment can significantly improve 3D lane detection. While the BEV branch still relies on IPM, future work may explore deformable attention mechanisms to further enhance robustness to complex road topographies. Additionally, extending this view-agnostic framework to multi-camera surround-view setups remains a promising direction for future research.
Summary updated