Scholarlink

<p>Curb is a key component of structured roads, making curb detection an important task in environmental perception for autonomous driving. To accurately extract curb edges around self-driving vehicles across multiple traffic scenarios, this paper proposes a novel road edge detection method based on 3D LIDAR. We design a road-edge point extraction algorithm that integrates three spatial features (horizontal distance <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\:{{\updelta\:}}_{\text{x}\text{y},\text{l}}\)</EquationSource> </InlineEquation>, height difference <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\:{{\updelta\:}}_{\text{z},\text{l}}\)</EquationSource> </InlineEquation>, angle difference Δθ) with loose thresholds, addressing the limitation of traditional fixed-threshold methods (e.g., Kang’s Hough transform, Peterson’s wavelet transform) which suffer over 30% missed detections on curved/sloped roads.By processing each LIDAR scan line independently and using the sign of Δθ (Δθ &lt; 0 for left curbs, Δθ &gt; 0 for right curbs) for left-right discrimination, we overcome the inefficiency of Wang’s clustering-based classification (computational complexity O(nlogn) vs. O(n)of our method, with ~ 67% more time consumption on a 4-core i7 CPU) and the failure of Zhang’s coordinate-dependent method on curved roads. Additionally, a two-stage cascade filter (DBSCAN + RANSAC) is designed for candidate points: DBSCAN eliminates discrete noise overlooked by Zhang’s sliding window, while RANSAC fits straight/curved curbs to remove obstacle-induced false detections (absent in Kang’s and Peterson’s methods), ensuring robustness in complex scenarios.Comprehensive experiments and comparisons with the deep learning method CurbNet show that our algorithm outperforms CurbNet in Precision (0.8674 vs. 0.8178), Recall (0.8704 vs. 0.8395), and F1 Score (0.8840 vs. 0.8325). It effectively detects curbs on straight/curved roads with obstacles (e.g., vehicles, pedestrians), with average Precision, Recall, and F1 Score above 80% across all test scenarios—verifying its accuracy and robustness. Notably, CurbNet requires an NVIDIA RTX 3090 GPU to reach 15&#xa0;Hz, while our lightweight design achieves 13&#xa0;Hz on a 4-core i7 CPU and needs no training (avoiding reliance on large labeled datasets like 3D-Curb), offering distinct engineering deployment advantages.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Robust Curb Detection Algorithm Based on LIDAR Sensor

  • 期刊论文

  • 发表日期：

    2026年1月4日

查看全文
  • Yongjian Zhu,
  • Zhang Qu,
  • Zihao Ding,
  • Shixuan Tian,
  • Fenqiang Chen,
  • Jiahui Lu

摘要

Curb is a key component of structured roads, making curb detection an important task in environmental perception for autonomous driving. To accurately extract curb edges around self-driving vehicles across multiple traffic scenarios, this paper proposes a novel road edge detection method based on 3D LIDAR. We design a road-edge point extraction algorithm that integrates three spatial features (horizontal distance \(\:{{\updelta\:}}_{\text{x}\text{y},\text{l}}\) , height difference \(\:{{\updelta\:}}_{\text{z},\text{l}}\) , angle difference Δθ) with loose thresholds, addressing the limitation of traditional fixed-threshold methods (e.g., Kang’s Hough transform, Peterson’s wavelet transform) which suffer over 30% missed detections on curved/sloped roads.By processing each LIDAR scan line independently and using the sign of Δθ (Δθ < 0 for left curbs, Δθ > 0 for right curbs) for left-right discrimination, we overcome the inefficiency of Wang’s clustering-based classification (computational complexity O(nlogn) vs. O(n)of our method, with ~ 67% more time consumption on a 4-core i7 CPU) and the failure of Zhang’s coordinate-dependent method on curved roads. Additionally, a two-stage cascade filter (DBSCAN + RANSAC) is designed for candidate points: DBSCAN eliminates discrete noise overlooked by Zhang’s sliding window, while RANSAC fits straight/curved curbs to remove obstacle-induced false detections (absent in Kang’s and Peterson’s methods), ensuring robustness in complex scenarios.Comprehensive experiments and comparisons with the deep learning method CurbNet show that our algorithm outperforms CurbNet in Precision (0.8674 vs. 0.8178), Recall (0.8704 vs. 0.8395), and F1 Score (0.8840 vs. 0.8325). It effectively detects curbs on straight/curved roads with obstacles (e.g., vehicles, pedestrians), with average Precision, Recall, and F1 Score above 80% across all test scenarios—verifying its accuracy and robustness. Notably, CurbNet requires an NVIDIA RTX 3090 GPU to reach 15 Hz, while our lightweight design achieves 13 Hz on a 4-core i7 CPU and needs no training (avoiding reliance on large labeled datasets like 3D-Curb), offering distinct engineering deployment advantages.