Abstract
In autonomous driving, detecting reliable and accurate lane marker positions is a crucial yet challenging task. Conventional approaches perform a pixel-level dense prediction task followed by sophisticated post-processing. We propose a method performing direct lane marker vertex prediction in an end-to-end manner, without any post-processing. Specifically, we translate the lane marker detection problem into a row-wise classification task. To compactly extract information about lane markers, we devise a novel Horizontal Reduction Module (HRM) to successively compress horizontal components, enabling final lane marker positions to be obtained via simple argmax operations. Experimental results on TuSimple and CULane benchmarks demonstrate that our method is on par with or outperforms state-of-the-art methods.
Introduction
Traditional lane detection systems typically rely on a two-stage pipeline: semantic segmentation followed by complex post-processing. In the first stage, a neural network predicts a pixel-level mask. In the second stage, algorithms like RANSAC or spline fitting are used to extract line coordinates from these masks. This approach has significant drawbacks: segmentation treats pixels independently, ignoring the global structure of lanes, and post-processing introduces a computational bottleneck that is difficult to tune across varying environments.
This paper introduces E2E-LMD, a framework that bypasses dense prediction by treating lane detection as a row-wise classification problem. By predicting lane coordinates directly, the model eliminates the need for heuristic-heavy post-processing.
Methodology: The E2E-LMD Architecture
The core innovation of E2E-LMD is framing lane detection as identifying the horizontal position of a lane marker within each row of an image.
1. Horizontal Reduction Module (HRM)
To transform a 2D feature map into a set of row-wise coordinates, the authors developed the Horizontal Reduction Module (HRM). Unlike standard pooling which discards information, the HRM uses a "Horizontal Pixel Unshuffle" layer. This operation rearranges spatial data into the channel dimension, preserving high-frequency information while reducing horizontal resolution.
2. Row-wise Classification
As shown in the architecture schema, the network uses successive HRMs to squeeze the feature map width until it reaches a single pixel. At this stage, each row is represented by a feature vector. The final layers perform classification across the original image width to determine the $x$-coordinate for each row. This architecture inherently respects the "one-point-per-row" geometric prior of lane markers.
Experimental Results
The proposed method was evaluated on two major benchmarks:
- TuSimple: Focused on highway scenarios. E2E-LMD achieved 96.22% accuracy with a remarkably low False Positive (FP) rate of 0.0308.
- CULane: A large-scale dataset with challenging urban and night scenes. Using an ERFNet backbone, the model achieved an F1-measure of 74.0, outperforming segmentation-based models like SCNN.
Discussion and Learned Representations
A key insight of this work is how the network learns to focus on lane structures through successive horizontal compression. By visualizing the feature maps at different stages of the reduction process, we can see the model progressively isolating lane markers from the background.
The inclusion of Squeeze-and-Excitation (SE) blocks within the HRM was found to be critical. These blocks allow the model to re-weight channels after the unshuffle operation, focusing on the most relevant features for coordinate prediction.
Conclusion
E2E-LMD demonstrates that row-wise classification is a powerful alternative to dense segmentation for lane detection. By leveraging the innate geometry of lanes and preserving spatial information through horizontal unshuffle operations, the model achieves state-of-the-art performance with a simpler, faster, and truly end-to-end pipeline.
Summary updated