军事知识服务平台
基于混合式点云数据训练的道路语义分割与几何信息自动提取
Road semantic segmentation and geometric information extraction based on hybrid point cloud data training
摘要 [目标]获取高精度道路几何信息是构建数字化道路模型和开展智能交通应用的基础.道路点云语义分割是几何信息提取的关键,但真实点云数据标注成本高,公开数据集规模有限,难以满足深度学习对多样化数据的需求.为此,提出一种真实与合成数据结合的混合训练策略,构建高精度分割模型,并实现高速公路关键几何参数自动提取.[方法]首先选取 SemanticKITTI 高速公路场景为真实数据集,基于 CARLA 仿真平台生成相似场景的合成点云.采用 DBSCAN 聚类剔除对向车道点云干扰,基于车辆轨迹线构建局部坐标系,逐点提取横截面点云.其次采用PointNet++进行语义分割训练,构建真实数据、小样本和混合数据 3 种训练模型,并通过平均交并比(mIoU)和总体准确度评价分割效果.最后,基于分割后的道路点云,提出几何参数自动提取框架,计算道路宽度、曲率、纵坡、横坡等指标,并与手动测量对比验证.[结果]真实数据训练模型 mIoU 为 61.5%,小样本训练仅 47%.混合数据训练后,mIoU 提升至 62.2%,对稀疏目标识别能力显著提升.几何信息提取中,道路宽度、纵坡和横坡的平均相对误差分别为 0.042,0.087,0.065,均达到较高精度.[结论]提出的混合训练策略有效提升点云语义分割精度,尤其改善了小样本条件下对稀疏目标的识别能力.基于分割点云的几何信息提取方法可准确获取道路关键参数,为高速公路数字化建模和智能交通应用提供了可靠技术支持.
Abstract:
[Objective]Obtaining high-precision road geometric information is the foundation for building digital road models and enabling intelligent transportation applications.Semantic segmentation of road point clouds is a critical step in geometric information extraction.However,the high cost of manual annotation for real point cloud data,along with the limited scale of existing public datasets,makes it difficult to meet the demand for diverse training data in deep learning.Therefore,this study proposes a hybrid training strategy that combines real and synthetic data to build a high-precision segmentation model and further achieves the automated extraction of key geometric parameters of expressways.[Method]First,this study selected expressway scene data from dataset SemanticKITTI as the real data,and generated synthetic point cloud data with similar road characteristics through CARLA simulation platform.DBSCAN clustering was applied to remove the interfering point clouds from opposite lanes.Local coordinate systems were constructed to extract cross-sectional point clouds based on vehicle trajectory lines.Subsequently,PointNet++network was used for semantic segmentation training and building three models,i.e.,real data model,small-sample model,and hybrid data model.The segmentation performance was evaluated by using the mean intersection over union(mIoU)and overall accuracy.Finally,an automatic geometric parameter extraction framework was proposed based on the segmented road point clouds.It calculated the key indicators,e.g.,road width,curvature,longitudinal slope,and cross slope,which were validated by comparison with manual measurements.[Result]The real data training model achieves the mIoU of 61.5%,while the small-sample model only reaches 47%.After hybrid data training,the mIoU improves to 62.2%,with particularly significant enhancement in recognizing sparse objects.In geometric information extraction,the average relative errors of road width,longitudinal slope,and cross slope are 0.042,0.087,and 0.065 respectively,all indicating high accuracy.[Conclusion]The proposed hybrid training strategy effectively improves the accuracy of point cloud semantic segmentation,especially enhancing the recognition of sparse objects under small-sample conditions.The geometric information extraction method based on segmented point clouds accurately obtains key road parameters,providing reliable technical support for highway digital modeling and intelligent transportation applications.
刊 名 公路交通科技 ISTICPKU
年,卷(期): 2026, 43(9)
分类号: U495
机标分类号: TP391 TP75 F301.24
在线出版日期: 2026年10月22日