ABSTRACT

Gravel roads management systems (GRMS) are in need of an integrated and cost-effective approach for condition data collection. In order to fulfil this need, this paper validates the practicality of utilising deep learning and image classifiers in collecting corrugation data from gravel roads. The used image classifier in this study was developed using the TensorFlow framework. This classifier has the capability to recognise and classify the corrugation severity on gravel roads into five levels. Furthermore, a pilot study was carried out in Laramie County, Wyoming to validate the applicability of the developed classifier in real practice. Three thousand images of gravel roads were captured from Laramie County gravel roads. Each captured image represents one gravel road section. The corrugation in the tested sections was evaluated by two methods, visual inspection and the developed image classifier. A confusion matrix was developed to determine the achieved accuracy by utilising the gravel roads corrugation image classifier. The confusion matrix showed that the developed image classifier has an 83% accuracy level in the practical field. The achieved accuracy level is considered sufficient for the purpose of GRMS.

摘要

碎石路管理系统 (GRMS) 需要一种综合且经济高效的路况数据收集方法。为了满足这一需求，本文验证了利用深度学习和图像分类器从碎石路收集波纹数据的实用性。本研究中使用的图像分类器是使用 TensorFlow 框架开发的。该分类器具有识别碎石路面的波纹严重程度并将其分为五个级别的能力。此外，在怀俄明州拉勒米县进行了一项试点研究，以验证开发的分类器在实际实践中的适用性。从拉勒米县碎石路拍摄了三千张碎石路图像。每个捕获的图像代表一个碎石路段。测试部分的波纹通过两种方法进行评估，目视检查和开发的图像分类器。利用砾石路波纹图像分类器开发了一个混淆矩阵来确定达到的精度。混淆矩阵表明，所开发的图像分类器在实际领域具有 83% 的准确度水平。达到的准确度水平被认为足以满足 GRMS 的目的。