Abstract
Lane detection is a fundamental component of autonomous driving systems. This paper introduces PriorLane, a transformer-only framework that achieves state-of-the-art (SOTA) performance on CULane and TuSimple benchmarks. Beyond standard visual features, PriorLane incorporates low-cost local prior knowledge (such as road maps) to enhance segmentation performance. A novel Knowledge Embedding Alignment (KEA) module is utilized to align these knowledge embeddings with image features. Extensive experiments on the Zjlab dataset demonstrate that PriorLane outperforms existing SOTA methods by 2.82% mIoU when prior knowledge is integrated.
Introduction
Lane detection remains a challenging "thin and long" problem in computer vision. Traditional Convolutional Neural Networks (CNNs) often struggle to capture the long-range dependencies necessary to trace lanes that disappear into the horizon or are obscured by obstacles. While existing methods attempt to use "shape priors," they frequently fail under severe occlusions, visual noise (like rain or glare), and complex geometries like intersections.
PriorLane addresses these issues by leveraging "prior knowledge"—low-cost map data that provides a blueprint of where the road should be, regardless of current visual conditions. By treating this map data as explicit prompts within a transformer architecture, the model can effectively "fill in the blanks" when visual cues are missing.
Methodology
The PriorLane architecture consists of three primary components:
- Hierarchical Backbone (MiT): The model employs a Mixed Transformer (MiT-B5) as an encoder. MiT uses overlapped patch merging to maintain local structural integrity while capturing the global context essential for segmenting thin lane lines.
- Knowledge Embedding Alignment (KEA): Prior knowledge, represented as Bird’s Eye View (BEV) grid maps, is often misaligned with the camera perspective due to coarse GPS data. The KEA module utilizes a Spatial Transformer Network (STN) and Active Rotation Filters (ARFs) to rotate and shift the map data, ensuring precise alignment with image features.
- Fusion Transformer (FT): Instead of simple concatenation, the model treats map data as explicit prompts. A dedicated Fusion Transformer uses self-attention to merge image features with knowledge embeddings, allowing the model to prioritize map data in regions where visual confidence is low.
Experimental Results
The researchers evaluated PriorLane on CULane, TuSimple, and a new custom dataset, Zjlab. The Zjlab dataset includes 869 frames with corresponding BEV grid maps to facilitate the training of the PriorLane framework.
Quantitative Performance
On the CULane dataset, the transformer-only baseline (MiT-Lane) achieved a 76.27 F1-measure, significantly outperforming established CNN-based models like RESA (74.04) and SCNN (73.58). On the Zjlab dataset, the full PriorLane-KEA model achieved 73.78% mIoU, a 2.82% improvement over the SOTA without prior knowledge.
Qualitative Robustness
Qualitative analysis reveals that PriorLane excels in scenarios where traditional models fail. It accurately predicts dotted lines and lanes hidden behind vehicles, and it can effectively distinguish between standard lane markings and stop lines at complex intersections.
Conclusion and Future Work
PriorLane demonstrates that pure transformer architectures, when augmented with aligned prior knowledge, provide a robust solution for lane detection. While the system currently requires meter-level positioning, future work will focus on handling scenarios with zero GPS signal and expanding the types of prior knowledge used (e.g., traffic signs and elevation maps) to further harden the system against extreme environmental conditions.
Summary updated