顾及拓扑结构的道路标线提取

Road markings extraction considering topological structure

道路标线是重要的交通标志信息,车载激光雷达点云为道路标线的提取提供了高精度的三维坐标和反射强度信息。由于扫描距离、目标材质等因素影响,不同目标会表现为相近的强度值,给道路标线的提取带来干扰;道路使用过程中的磨损老化会破坏标线原有的结构,造成标线提取后出现间断不连续问题;此外,道路标线种类多样且实际中出现的概率不同,导致分割网络提取结果中样本较少的类别分类精度较低。为此,本文提出了一种可准确提取各类标线并具有拓扑稳健性的分割+分类两阶段提取方法。首先,使用多层感知机自适应学习强度与其影响因素之间的关系,对路面点云进行强度改正;然后,提出链式空间拓扑网络LST-Net对道路上的所有标线进行语义分割,引入行列卷积、注意力机制捕捉标线结构信息,加入拓扑惩罚对其训练,确定标线的位置;最后,使用YOLOv5对标线进行检测,单独训练分类网络,解决分割中样本不均衡的问题,完成对标线的分类。在不同场景的3份车载点云上进行试验,结果表明本文方法标线提取精度达到94.1%,召回率达到95.6%,具有较强的实用性和有效性。

Road markings are important traffic sign information,and onboard LiDAR point clouds provide high precision 3D coordinates and reflectance intensity information for their extraction.Due to factors such as scanning distance and target material,the different object may exhibit similar intensity values,causing interference in the extraction of road markings.Wear and aging during road use can also damage the original structure of the markings,resulting in discontinuities after extraction.In addition,the diversity of road markings and their varying occurrence frequencies in practice can lead to low classification accuracy for categories with fewer samples in the segmentation network extraction results.To address these issues,this paper proposes a two-stage segmentation and classification extraction method that accurately extracts various types of markings and has topological robustness.Firstly,a multi-layer perceptron is used to adaptively learn the relationship between intensity and its influencing factors,and to perform intensity correction on the road point clouds.Secondly,the semantic segmentation network link spatial topology net(LST-Net)is proposed to segment all road markings,which captures line structure information using row-column convolution and attention mechanisms,and is trained with topological punishment to determine the positions of markings.Finally,YOLOv5 is used to detect the markings,and a separate classification network is trained to address the issue of sample imbalance in segmentation.Experiments are conducted on three sets of point clouds from different driving scenarios,and the results show that our approach achieves a marking extraction accuracy of 94.1%and a recall rate of 95.6%,demonstrating strong practicality and effectiveness.