01763562

Component

https://annualmeeting.myt...nlineProgram/Details/15692

The introduction of Tunnel Monitoring System(TMS) in 2011, it has so far found many defects in the tracks, roadbeds, and urban railroad tunnels and has made many achievements in taking a proactive step. However, the video is still being viewed and judged by a person with the train control center. If the filmed images can be analyzed with big data to determine if there are any abnormalities in the tunnel, the accuracy of the judgment and human error can be prevented by combining mechanical and human analysis, and automation will be possible in the long term. In this paper, the method of classifying video images in the tunnel was attempted by utilizing the transfer learning of CNN. The results showed that among CNN's three models (AlexNet, VGGNet, GoogLeNet), the Fine-Tuning model has the highest accuracy and lowest loss rate. The evaluation results showed that the accuracy of normal and abnormal images differed depending on the training data volume. The model to performance of expected to improve significantly under actual conditions if sufficient video data can be obtained in the future, although it has been more than 99% accurate only in limited conditions that only railway signs and railway lines are available.

This paper was sponsored by TRB committee AKB60 Standing Committee on Tunnels and Underground Structures.

TRBAM-21-00692

English

12p

2021

Transportation Research Board 100th Annual Meeting

Web

Figures; References; Tables

Classification; Imagery; Machine learning; Machine vision; Monitoring; Neural networks; Subways; Tunnels; Urban transit

Bridges and other structures; Data and Information Technology; Railroads

Transportation Research Board Annual Meeting 2021 Paper #TRBAM-21-00692

TRIS, TRB, ATRI

Dec 23 2020 11:06AM