CANet: Curved Guide Line Network with Adaptive Decoder for Lane Detection

Zhongyu Yang, Chen Shen, Wei Shao, Tengfei Xing, Runbo Hu, Pengfei Xu, Hua Chai, Ruini Xue

TL;DR

This paper introduces CANet, a novel deep learning approach for lane detection that tackles the challenge of "corner lanes" by employing a U-shaped "curved guide line" to improve feature aggregation. CANet also features an adaptive decoder that intelligently switches between row-wise and column-wise classification, leading to more consistent lane predictions and achieving state-of-the-art results on multiple benchmarks.

What Problem Does This Paper Address?

Accurate lane detection is a critical component for advanced driver-assistance systems (ADAS) and autonomous driving. However, existing methods struggle with complex road scenarios, particularly with lanes that have sharp turns or "corner lanes" where the grazing angle is small. Traditional approaches often rely on rectangular guides or fixed classification strategies, which can lead to missed detections or inaccurate lane shapes, especially for horizontal or near-horizontal lanes. This paper aims to improve lane detection robustness and accuracy, particularly in challenging corner scenarios, by introducing a more flexible and adaptive network architecture.

Methods

CANet is a top-down deep learning lane detection network with several key innovations:

  • Curved Guide Line: Instead of a traditional rectangular guide, CANet uses a U-shaped "curved guide line" (partially an ellipse) to define the initial lane origins. This design increases the effective grazing angles for corner lanes, making them easier for the network to detect and recall.
  • Heatmap-based Instance Aggregation: Lane instances are initially detected via a heatmap response on the curved guide line. This allows for the aggregation of relevant lane features at a global semantic level.
  • Conditional Convolution: The network uses conditional convolution to refine the heatmap response across the entire lane, leveraging the aggregated features.
  • Adaptive Decoder: A novel adaptive decoder mechanism is employed. It uses a Gaussian mask to supervise the lane generation process. Based on the shape of this mask, the decoder can dynamically choose between row-wise or column-wise classification during inference. This adaptability allows it to handle lanes with varying orientations more effectively.
  • Key Point Capturing with Normalized Gaussian Kernel: To mitigate position inaccuracies caused by downsampling, CANet uses a normalized Gaussian kernel with an offset as supervision for key point detection.

Datasets

CANet was evaluated on three widely used lane detection datasets:

  • CULane: A challenging dataset known for its diverse and complex road scenarios.
  • CurveLanes: Specifically designed to include more difficult curved lane scenarios.
  • TuSimple: A simpler dataset often used for benchmarking.

Key Insights and Ideas

  • The Geometry of Detection Matters: The paper highlights how the choice of a "guide line" significantly impacts the network's ability to detect lanes, especially those with small grazing angles. A curved guide line is demonstrably better than a rectangular one for improving recall in these scenarios.
  • Adaptability is Key for Diverse Lanes: The insight that lanes can be horizontal, vertical, or curved necessitates a flexible detection strategy. CANet's adaptive decoder, which can switch between row-wise and column-wise classification, addresses this by dynamically adjusting to the lane's shape.
  • Unified Prediction for Range and Location: By using the heatmap response range to determine the lane's extent, CANet avoids the need for separate range classifiers, which can lead to inconsistencies. This unified approach simplifies the prediction process and improves reliability.

Key Results

CANet achieved state-of-the-art (SOTA) performance on the evaluated datasets:

  • CULane: Achieved an overall F1-score of 79.86%, surpassing the previous best of 79.63%. It also showed strong performance on difficult subsets like Crowded, Dazzle, and Night.
  • CurveLanes: CANET-L reached an SOTA F1-score of 87.87%, a 1.7 percentage point improvement. The smaller CANET-S model also achieved competitive results with significantly lower computational cost.
  • TuSimple: While all models performed well, CANet provided marginal improvements across most indicators.
  • Effectiveness of Curved Guide Line: Table 3 shows that the curved guide line significantly improved recall across all grazing angle groups, particularly for the smallest group (0°-30°), demonstrating its effectiveness in handling corner lanes.

Limitations and Future Work

  • Computational Cost: While CANet-S offers a good trade-off, the larger CANET-L model's computational cost is not explicitly detailed as a limitation, but it's implied by the comparison with CANET-S.
  • Future Work: The paper focuses on improving lane detection accuracy and robustness. Future work could explore real-time performance optimizations, integration with other driving tasks, or generalization to even more extreme scenarios. The adaptive decoder mechanism itself could be further refined or explored in other computer vision tasks.

Summary updated

Abstract

Lane detection is challenging due to the complicated onroad scenarios and line deformation from different camera perspectives. Lots of solutions were proposed, but can not deal with "corner lanes" well. To address this problem, this paper proposes a new top-down deep learning lane detection approach, CANet. A lane instance is first responded by the heatmap on the U-shaped "curved guide line" at global semantic level, thus the corresponding features of each lane are aggregated at the response point. Then CANet obtains the heatmap response of the entire lane through conditional convolution, and finally decodes the point set to describe lanes via adaptive decoder. The prototype is implemented with Pytorch, and evaluated against 3 well-known datasets extensively. The experimental results show that CANet reaches SOTA in different metrics.
BibTeX
@misc{yang2023canet,
  author = {Zhongyu Yang and Chen Shen and Wei Shao and Tengfei Xing and Runbo Hu and Pengfei Xu and Hua Chai and Ruini Xue},
  title = {{CANet: Curved Guide Line Network with Adaptive Decoder for Lane Detection}},
  year = {2023},
  eprint = {2304.11546},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2304.11546},
}

Loading related content...

Figure