Abstract
Lane detection is a fundamental task in autonomous driving, yet it remains challenging in complex scenarios such as night driving, occlusions, and unmarked roads. We propose the Hybrid Global-Local Perception Network (HGLNet), a novel architecture that addresses these issues by combining a sparse, adaptive global extraction head with a multi-hierarchy local feature aggregator. HGLNet effectively balances the need for global context to infer lane existence and local features for precise boundary localization. Our model achieves state-of-the-art results on major benchmarks, including CULane and LLAMAS, while operating at a real-time speed of 104 FPS.
Introduction
Reliable lane detection is essential for safe autonomous navigation. However, standard computer vision models often struggle with "hard cases" where visual evidence is sparse or absent—such as faded markings, heavy traffic occlusion, or low-light conditions. The core challenge lies in a fundamental tension: globality is required to infer lane positions from the overall scene context, while locality is necessary to precisely define lane boundaries using fine-grained textures. Existing methods often fail to integrate these two aspects efficiently, leading to either poor localization or detection failures in cluttered environments.
Methods: The HGLNet Architecture
HGLNet introduces a dual-pathway approach to bridge the gap between global reasoning and local precision.
Global Extraction Head
Inspired by Deformable DETR, the Global Extraction Head avoids the high computational cost of dense Transformers. Instead of attending to every pixel, it utilizes adaptive sampling points around the lane. These points are processed through self-attention layers to capture long-range semantic context. This allows the model to "guess" lane positions even when they are occluded, by reasoning about the relative positions of other objects, such as the vehicle ahead.
Multi-hierarchy Feature Aggregator (MFA)
The MFA module focuses on capturing local consistency and fine-grained details. It employs Dilated Convolutions to expand the receptive field without increasing the parameter count, alongside a Context-aware Attention mechanism. This enables the network to learn "3D weights" (spatial and channel) specifically optimized for slender objects like lane markings. By aggregating features across different scales, the MFA ensures that the model remains robust to variations in lane width and curvature.
Hybrid Integration
The network processes global and local features in parallel and fuses them to produce final predictions. This hybrid architecture allows the model to simultaneously determine the presence of a lane (global) and accurately regress its coordinates (local).
Experiments and Results
We evaluated HGLNet on three industry-standard benchmarks: CULane, TuSimple, and LLAMAS.
Performance Benchmarks
HGLNet consistently outperforms previous state-of-the-art models, such as CLRNet and LaneATT, across multiple datasets. On the CULane dataset, which is known for its difficult scenarios like crowded roads and night driving, HGLNet achieved an 81.83% F1 score with a DLA-34 backbone. Notably, the model showed substantial gains in the "Shadow" and "Night" categories, where traditional methods often fail due to low contrast.
On the LLAMAS dataset, HGLNet established a new state-of-the-art with a 96.20% F1 score, surpassing previous benchmarks by a significant margin. This performance underscores the model's ability to handle high-speed highway environments with high precision.
Efficiency and Real-time Performance
A key advantage of HGLNet is its computational efficiency. By utilizing sparse adaptive sampling instead of dense global attention, the model avoids the quadratic complexity typically associated with Transformer-based architectures. The ResNet-18 variant achieves a throughput of 116 FPS, while the more robust DLA-34 version maintains 104 FPS. These speeds are well above the requirements for real-time processing in autonomous vehicles.
Conclusion
HGLNet demonstrates that the combination of sparse global attention and multi-hierarchy local feature aggregation is highly effective for lane detection. By decoupling the tasks of existence detection and precision localization, the model achieves superior accuracy and robustness in difficult real-world conditions. While "Cross" scenarios (intersections) remain a challenge for the field, HGLNet shows marked improvement. Future work will explore the application of the MFA module to other slender object detection tasks and further optimize the adaptive sampling strategy for complex topological structures.
Summary updated