Abstract: In autonomous driving, detecting reliable and accurate lane marker positions is a crucial yet challenging task. The conventional approaches for the lane marker detection problem perform a pixel-level dense prediction task followed by sophisticated post-processing that is inevitable since lane markers are typically represented by a collection of line segments without thickness. In this paper, we propose a method performing direct lane marker vertex prediction in an end-to-end manner, i.e., without any post-processing step that is required in the pixel-level dense prediction task. Specifically, we translate the lane marker detection problem into a row-wise classification task, which takes advantage of the innate shape of lane markers but, surprisingly, has not been explored well. In order to compactly extract sufficient information about lane markers which spread from the left to the right in an image, we devise a novel layer, which is utilized to successively compress horizontal components so enables an end-to-end lane marker detection system where the final lane marker positions are simply obtained via argmax operations in testing time. Experimental results demonstrate the effectiveness of the proposed method, which is on par or outperforms the state-of-the-art methods on two popular lane marker detection benchmarks, i.e., TuSimple and CULane.
Introduction
Lane detection is a fundamental requirement for autonomous driving, yet it remains a complex challenge. Most conventional deep learning approaches treat it as a semantic segmentation task, where the network labels every pixel as "lane" or "not lane." However, because lane markers are thin, structured lines, pixel-level masks are often noisy or disconnected. This necessitates sophisticated, hand-tuned post-processing steps—such as curve fitting or clustering—which are computationally expensive and difficult to integrate into a single training loop. This paper introduces the E2E-LMD (End-to-End Lane Marker Detection) framework, which reimagines lane detection as a clean row-wise classification task, predicting lane coordinates directly in one go.
Methodology
The core innovation of E2E-LMD is translating lane detection into row-wise classification. Instead of binary pixel labeling, the model determines which horizontal position (column) contains the lane marker for every specific row.
Architecture and the HRM
The architecture consists of three main stages:
- Encoder-Decoder Backbone: A standard backbone (e.g., ResNet or ERFNet) extracts initial feature maps.
- Horizontal Reduction Modules (HRMs): To transition from 2D images to 1D row-wise representations, the authors developed HRMs. These layers utilize a "horizontal pixel unshuffle"—the inverse of the pixel shuffle used in super-resolution—to move horizontal spatial information into the channel dimension. This allows the network to compress the width of the feature map while preserving the fine-grained information necessary for precise localization.
- Dual Branches: The final stage splits into a Vertex Location Branch (predicting the $x$ position for every row $y$) and a Vertex Confidence Branch (determining if a lane exists at that row).
Key Insights
- Exploiting Geometric Priors: By processing images row-by-row, the model respects the naturally vertical or diagonal geometry of lane markers, simplifying the learning task compared to general segmentation.
- Spatial-to-Channel Shift: The pixel unshuffle technique "repackages" spatial pixels into channels rather than discarding data via pooling. This enables subsequent layers to analyze the entire road width within a single feature vector.
- Feature Recalibration: The inclusion of Squeeze-and-Excitation (SE) blocks after horizontal compression allows the model to recalibrate which features are most important once spatial dimensions have been squeezed.
Experimental Results
The E2E-LMD framework was evaluated on two industry-standard benchmarks, demonstrating both accuracy and robustness:
- TuSimple: Using a ResNet-34 backbone, the model achieved 96.22% accuracy. It notably produced significantly fewer False Positives (FP) than previous state-of-the-art models like SCNN.
- CULane: With an ERFNet backbone, the model achieved an F1-measure of 74.0, outperforming established models like SCNN (71.6) and SAD (71.8).
Qualitative analysis shows that the model excels in difficult conditions, such as tunnels or heavy shadows, where traditional segmentation often fails. At inference time, lane coordinates are obtained via a simple argmax operation, eliminating the need for complex curve fitting.
Conclusion
E2E-LMD provides a streamlined, effective alternative to segmentation-based lane detection. While highly robust, the model faces challenges in cases of severe curves or extreme occlusions where row-wise assumptions are strained. Future work may involve using Neural Architecture Search (NAS) to further optimize the Horizontal Reduction Module for real-time deployment on low-power automotive hardware.
Summary updated