Automatically detect and classify asphalt pavement raveling severity using 3D technology and machine learning
Automatically detect and classify asphalt pavement raveling severity using 3D technology and machine learning
复制标题
使用 3D 技术和机器学习自动检测和分类沥青路面松散严重程度
DOI：
--
复制
发表时间：
2020
影响因子：
--
通讯作者：
A. Chatterjee
中科院分区：
文献类型：
--
作者：
Y. Tsai;Yipu Zhao;Bruno Pop;A. Chatterjee
松脱是美国高速公路路面上最常见的沥青路面病害之一。打散会导致安全问题，如松散的石头和打滑；驾驶质量差，路面/轮胎噪音大；并缩短路面寿命。传统的公路勘测方法采用人工目测，耗时长、主观，对公路工作人员危害大。该研究项目由国家合作公路研究计划（NCHRP）创新值得探索分析（IDEA）项目竞争性选择和赞助，本研究的目的是利用3D路面数据开发一种准确的行驶检测和分类算法，该算法已成为美国州交通部（DOTs）用于路面状况评估的主流技术。并根据实际运输机构的遇险协议（严重级别1、2和3），使用大规模的真实数据全面验证这些方法。在乔治亚州的I-85和I-285公路上，共收集了65英里的3d路面数据，用于培训和测试。开发了三种监督机器学习技术——adaboost与决策树、支持向量机（SVM）和随机森林——用于收集数据中的旅行检测和分类。随机森林分类器的性能最好，在真实世界大规模数据上，精度值从3级旅行的75.6%到0级（无）旅行的97.6%，召回值从1级旅行的86.9%到0级旅行的96.1%。该方法经过大规模验证和细化，已成功应用于整个乔治亚州州际公路系统1452.5英里的沥青路面调查中。该方法可应用于其他交通运输机构，更安全、更有效地评估道路行驶状况。
Raveling is one of the most common asphalt pavement distresses that occur on US highway pavements. Raveling results in safety concerns such as loose stones and hydroplaning; poor ride quality and road/tire noise; and shortened pavement longevity. Traditional raveling survey methods involve manual visual inspection, which is time consuming, subjective, and hazardous to highway workers. With the research project competitively selected and sponsored by the National Cooperative Highway Research Program (NCHRP) Innovation Deserving Exploratory Analysis (IDEA) program, the objective of this study is to develop an accurate raveling detection and classification algorithm using 3D pavement data that has become mainstream technologies for state Department of Transportations (DOTs) in the US for pavement condition evaluation, and to comprehensively validate these methods using large-scale, real-world data based on actual transportation agencies’ distress protocol (Severity levels 1, 2, and 3). A total of 65 miles of 3 D pavement data was collected on I-85 and I-285 in Georgia for training and testing. Three supervised machine learning techniques —AdaBoost with decision trees, support vector machine (SVM) and random forests—were developed for the detection and classification of raveling in the collected data. The random forest classifier had the b est performance, with precision values ranging from 75.6% for level 3 raveling to 97.6% for level 0 (no) raveling and recall values ranging from 86.9% for level 1 raveling to 96.1% for level 0 raveling on real world large-scale data. The developed raveling detection and severity level classification method has been successfully implemented to entire Georgia’s interstate highway system with1452.5 survey miles of asphalt pavements after the large-scale validation and refinement. The proposed method for raveling detection can be deployed to other transportation agencies for safer and more efficient assessment of roadway raveling conditions.