汽车工程 ›› 2025, Vol. 47 ›› Issue (6): 1133-1143.doi: 10.19562/j.chinasae.qcgc.2025.06.012

• • 上一篇    

基于LiDAR点云特征补全的雪天无人车目标检测

朱凌云(),王海洋   

  1. 重庆理工大学计算机科学与工程学院,重庆 400054
  • 收稿日期:2024-11-20 修回日期:2025-02-04 出版日期:2025-06-25 发布日期:2025-06-20
  • 通讯作者: 朱凌云 E-mail:zhulingyun@cqut.edu.cn
  • 基金资助:
    第二十七届中国科协年会学术论文。重庆市教委科学技术研究项目(KJQN202001118);重庆市技术创新与应用发展重点项目(cstc2021jscx-dxwtBX0018);重庆市自然科学基金(CSTB2022NSCQ-MSX0493)

Autonomous Vehicle Object Detection by LiDAR Point Cloud Feature Completion in Snowfall Scenarios

Lingyun Zhu(),Haiyang Wang   

  1. School of Computer Science and Engineering,Chongqing University of Technology,Chongqing 400054
  • Received:2024-11-20 Revised:2025-02-04 Online:2025-06-25 Published:2025-06-20
  • Contact: Lingyun Zhu E-mail:zhulingyun@cqut.edu.cn

摘要:

在降雪气候条件下,雪花颗粒对激光雷达的干扰会导致点云特征缺失,严重影响LiDAR三维目标检测模型的准确性。本文提出一种基于Transformer架构的雪天点云特征补全检测算法:首先设计点云损失补全模块,通过多头注意力机制与混合密度网络联合提取原始点云缺失特征;其次构建编码器-解码器结构实现缺失特征生成,并开发融合重定义模块通过通道注意力机制实现特征对齐;最后优化预测框输出策略提升检测可靠性。在CADC数据集上,汽车与行人检测精度分别提升2.06%和2.73%;在KITTI数据集上3类目标平均精度提升1.51%。通过量化分析降雪强度与点云生成数量的影响规律,验证了本文所提方法的鲁棒性和工程适用性。

关键词: 激光雷达, 降雪气候, 点云补全, 三维目标检测, 自动驾驶

Abstract:

Under snowy conditions, interference from snowflakes on LiDAR leads to point cloud feature loss, significantly degrading the accuracy of 3D object detection models. In this paper, a Transformer-based feature completion algorithm for snow-affected point clouds is proposed. Firstly, a point cloud loss completion module is designed to jointly extract missing features from raw point clouds using multi-head attention mechanisms and mixed density networks. Subsequently, an encoder-decoder architecture is constructed for feature generation, coupled with a fusion redefinition module that aligns features via channel attention mechanisms. Finally, the bounding box prediction strategy is optimized to enhance detection reliability. Experimental results demonstrate that the proposed method achieves improvement of 2.06% and 2.73% in car and pedestrian detection accuracy on the CADC dataset, and a 1.51% average precision gain across three object categories on the KITTI dataset. Quantitative analysis of snowfall intensity and point cloud generation patterns further validates the robustness and engineering applicability of the proposed method.

Key words: LiDAR, snowfall conditions, point cloud completion, 3D object detection, autonomous driving