Journal
INTERNATIONAL JOURNAL OF TRANSPORTATION SCIENCE AND TECHNOLOGY
Volume 11, Issue 2, Pages 298-309Publisher
KEAI PUBLISHING LTD
DOI: 10.1016/j.ijtst.2021.04.008
Keywords
Pavement distress type detection and; classification; Road maintenance operation; Camera-based pavement monitoring system; Convolutional neural network
Ask authors/readers for more resources
This study proposes and evaluates a methodology for automated detection and classification of pavement distress types using Convolutional Neural Networks (CNN) and a low-cost video data collection strategy. The trained CNN model achieves a detection rate of 83.8% and demonstrates high classification accuracy, providing an effective solution for road maintenance.
Pavement distresses, including cracking and disintegration, deteriorate road user's comfort, damage vehicles, increase evasive maneuvers, and increase emissions. Transportation agencies spend a significant portion of their budget to monitor and maintain road pavements. Pavement distress can be identified through manual surveys, i.e., visual inspections of pavement images captured by an inspection vehicle. To reduce manual inspection costs, research and industry have moved quietly towards the development and implementation of automated road surface monitoring systems. Considering the latest research developments, the objective of this work is to propose and evaluate a methodology for automated detection and classification of pavement distress types using Convolutional Neural Networks (CNN) and a low-cost video data collection strategy. In this work, pavement distress types are categorized as linear or longitudinal crack, network crack, fatigue crack or pothole, patch, and pavement marking. The models are trained and tested based on an image dataset collected from Montreal's road pavements. A sensitivity analysis is carried on for evaluating different regularization scenarios and data generation strategies especially scaling and partitioning the input image. The detection rate and classification accuracy of the proposed approach with the trained CNN model reaches 83.8% over the test set, which is promising compared with the literature. More specifically, the F1-scores for pothole, patch, marking, crack-linear and crack-network classes are 0.808, 0.802, 0.860, 0.796, and 0.813, respectively. However, by merging linear and network crack classes, the F1-score over the merged class increases to 0.916.& COPY; 2021 Tongji University and Tongji University Press. Publishing Services by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/ licenses/by-nc-nd/4.0/).
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available