Introduction
Lane detection is a fundamental component of Autonomous Driving Systems (ADS). However, real-world environments present significant challenges, including complex topologies (forks, merges, sharp curves) and frequent occlusions. Traditional methods often rely on rigid geometric priors or local keypoints, which struggle to adapt to these messy scenarios. This paper proposes a novel transformer-based architecture that generates dynamic convolutional kernels tailored for each lane line, leveraging global context to overcome these limitations.
Methodology
The proposed framework moves beyond local features by using a Transformer to capture the global structure of the road. The architecture consists of three main stages:
- Feature Extraction: A standard CNN backbone (e.g., ResNet) extracts base feature maps from the input image.
- Dynamic Kernel Head: This core component utilizes a Transformer Encoder to infuse global context into the feature map. Subsequently, a Transformer Decoder employs learnable "lane queries"—prototypes representing potential lane paths—to attend to these global features. These queries are transformed via Multi-Layer Perceptrons (MLPs) into unique convolutional kernels for each lane.
- Lane Detection Head: The generated kernels are convolved with the feature map to produce a probability heat map and a horizontal offset map for precise lane refinement.
The model uses bipartite matching via the Hungarian algorithm during training to match predictions with ground-truth lanes, simplifying the label assignment process.
Experimental Results
The model was evaluated on four major benchmarks: OpenLane, CurveLanes, CULane, and TuSimple. It achieved state-of-the-art (SOTA) performance across all datasets. Notably, on OpenLane, the model reached an F1 score of 63.40, a 4.30-point improvement over previous methods. On CurveLanes, it achieved an F1 score of 88.47.
Qualitative analysis demonstrates the model's ability to "hallucinate" occluded lane segments by relying on the global structure captured by the Transformer. Even in cases of heavy traffic or complex forks, the dynamic kernels maintain coherent lane tracks where previous local-feature-based models often failed. Efficiency is also a key highlight; a lightweight version of the head on a ResNet-18 backbone achieves 105 FPS, making it viable for real-time use.
Conclusion
By integrating Transformers into the dynamic kernel generation process, this research effectively addresses the long-standing issues of occlusion and complex topology in lane detection. The approach is both accurate and computationally efficient, providing a robust solution for real-time autonomous driving applications.
Summary updated