Source-linked AI summary
Focus on Local: Detecting Lane Marker from Bottom Up via Key Point
Zhan Qu, Huan Jin, Yang Zhou, Zhen Yang, Wei Zhang
TL;DR
Lane detection must capture precise local positions and complex global lane shapes, but mainstream segmentation and detection pipelines rely on high-dimensional outputs and post-processing. FOLOLane instead estimates and associates local key points with two network heads, then decodes them into global curves. It reports strong public-dataset results, including Acc: 96.92% on TuSimple and F1 score: 78.8% on CULane, while supporting architecture compatibility and real-time processing.
Problem
Existing lane-detection frameworks have difficulty representing curved lane lines with precise local positions and global shapes, while segmentation adds noisy redundancy and detection uses limited local information.
Method
FOLOLane models local lane geometry through key-point existence and offset heads, associates points belonging to the same lane, and decodes them bottom-up into global curves.
Results
Acc: 96.92% on TuSimple and F1 score: 78.8% on CULane are reported, with substantial improvement over existing methods and excellent generalization across two public datasets.
Takeaways & Limitations
The local bottom-up formulation reduces CNN output complexity while avoiding ambiguous and noisy pixels, supporting accuracy, generalization, and compatibility with evolving network architectures.
Takeaways & Limitations
Future work includes incorporating more powerful architectures, such as self-attention models, and applying FOLOLane to the MindSpore computing framework.
Abstract
from arXiv · showhide
Mainstream lane marker detection methods are implemented by predicting the overall structure and deriving parametric curves through post-processing. Complex lane line shapes require high-dimensional output of CNNs to model global structures, which further increases the demand for model capacity and training data. In contrast, the locality of a lane marker has finite geometric variations and spatial coverage. We propose a novel lane marker detection solution, FOLOLane, that focuses on modeling local patterns and achieving prediction of global structures in a bottom-up manner. Specifically, the CNN models lowcomplexity local patterns with two separate heads, the first one predicts the existence of key points, and the second refines the location of key points in the local range and correlates key points of the same lane line. The locality of the task is consistent with the limited FOV of the feature in CNN, which in turn leads to more stable training and better generalization. In addition, an efficiency-oriented decoding algorithm was proposed as well as a greedy one, which achieving 36% runtime gains at the cost of negligible performance degradation. Both of the two decoders integrated local information into the global geometry of lane markers. In the absence of a complex network architecture design, the proposed method greatly outperforms all existing methods on public datasets while achieving the best state-of-the-art results and real-time processing simultaneously.
1. Introduction
FOLOLane reframes lane detection as local key-point estimation and association, then integrates local geometry into global lane curves bottom-up. This design targets limitations of segmentation and detection pipelines while achieving strong accuracy and generalization.
- Lane detection supports autonomous-driving decisions, lane keeping, adaptive cruise control, and high-precision positioning against maps.
- Existing segmentation and detection methods struggle to represent curved lane lines while preserving precise local positions and global shapes.Segmentation adds clustering and curve fitting over redundant, noisy pixels; detection is insensitive to pixel-level errors for large lane markers.
- A fully convolutional network uses separate heads to estimate local key-point existence and refine key-point geometry, while decoders associate points into lane instances.The method focuses CNN capacity on a spatially limited local scope and uses local information to construct global curves.
- Acc: 96.92% on TuSimple and F1 score: 78.8% on CULane are reported alongside excellent generalization across two public datasets.
- FOLOLane formulates lane detection as multi-key-point estimation and association, inspired by bottom-up human pose estimation.The approach models local geometry and integrates it into global lane predictions.
2. Related Work
Prior lane-marker methods mainly use detection or segmentation, while FOLOLane connects lane detection with bottom-up key-point detection and association. It addresses redundancy and precision issues by predicting local points and their relationships.
- Lane Marker Detection: Lane-marker detection methods are broadly categorized into detection-based and segmentation-based approaches.
- Lane Marker Detection: Detection methods sample lane markers or predict lane existence and positions, whereas segmentation methods classify pixels and cluster them into lane instances.
- Bottom-Up Human Key Point Detection: Bottom-up human key-point methods detect points and associate them into individual instances using graphs, affinity fields, or embeddings.
- Bottom-Up Human Key Point Detection: FOLOLane abstracts lane detection as discrete key-point detection and association, replacing redundant pixel-wise joint-distribution prediction with a point-based formulation.
- Bottom-Up Human Key Point Detection: FOLOLane produces local-curve logits and uses a decoder to integrate them into global curve instances.
3. Methodology
FOLOLane decomposes lane detection into local keypoint estimation and association, then decodes these local predictions into global curve instances. Two decoder options trade precision for efficiency while integrating local geometry across neighboring rows.
- Local keypoint estimation: FOLOLane decomposes lane detection into estimating keypoint existence and offsets, followed by a global geometry decoder that generates curve instances.The fully convolutional network uses separate heads for local keypoint heatmaps and geometric offsets.
- Local keypoint estimation: The heatmap head models the probability that each pixel is a curve keypoint, while Gaussian-smoothed targets provide neighborhood supervision.When multiple keypoints assign a pixel, the maximum ground-truth value is retained.
- Local keypoint estimation: Penalty-reduced focal loss addresses sparse-keypoint class imbalance by reducing supervision from ambiguous negative pixels.Only pixels with ground truth equal to 1 are treated as positive; β and γ control ambiguity and simple-sample penalties.
- Local geometry construction: The geometry head predicts horizontal offsets to three neighboring keypoints separated by a fixed vertical interval, enabling local curve recovery.The offsets describe points above, horizontally adjacent to, and below the reference pixel.
- Global geometry decoding: Greedy decoding starts from a row with the most local maxima, iteratively refines points and explores neighbors, then groups associated points into curves.It uses a coarse-to-fine process that can produce precise geometry but has lower efficiency.
- Global geometry decoding: Efficient decoding finds all keypoint candidates initially and parallelizes refinement, neighbor prediction, and association construction before low-overhead grouping.The decoder uses index operations for the final grouping step and is designed to exploit device parallelism.
4. Experiments
Experiments evaluate FOLOLane on TuSimple and CULane, including comparisons with state-of-the-art methods, ablations of local designs and decoding, efficiency tests, and cross-dataset generalization.
- Quantitative results: FOLOLane achieves 78.8 F1 on the CULane testing set, establishing a new state-of-the-art result.
- Quantitative results: FOLOLane exceeds the second-best method on TuSimple by 0.17% despite near-saturated accuracy across methods.The dataset has a limited scale and homogeneous highway scenario.
- Ablation study: Keypoint estimation raises CULane F1 from 74.2 to 76.6 by suppressing ambiguous and noisy pixels.
- Ablation study: Coarse-to-fine refinement improves inference F1 by 0.9%, while extending it to training improves form sampling by 1.3%.The strategy guides spatially related representations toward more accurate curve geometry.
- Ablation study: Parallel decoding gains 16 ms, or 64% runtime, over greedy decoding at a cost of 0.8% performance degradation.The efficiency-oriented decoder replaces iterative neighboring-keypoint decoding with parallel processing.
- Efficiency-oriented implementation: Replacing ERFNet with BiSeNet yields more than 100 fps and a 77.5 F1 measure, while retaining compatibility with alternative network architectures.
- Generalization: A CULane-trained FOLOLane checkpoint reaches 84.36% on TuSimple, surpassing other methods by nearly 20% in cross-dataset generalization.The experiment evaluates inference on TuSimple without retraining from a CULane-trained checkpoint.
5. Conclusion and Future Work
FOLOLane uses a local, bottom-up lane-detection strategy that reduces CNN output dimensionality and improves accuracy and generalization. Future work includes integrating more powerful architectures and using MindSpore.
- FOLOLane’s keypoint estimation and coarse-to-fine refinement avoid ambiguous and noisy pixels, improving lane-curve geometry accuracy.
- The local-geometry, bottom-up pipeline simplifies lane detection by reducing CNN output dimensionality, supporting the reported performance and generalization.
- Future work will incorporate architectures with self-attention and apply FOLOLane to the MindSpore deep-learning framework.