Logo TU Ilmenau


Univ.-Prof. Dr.-Ing. Horst-Michael Groß

Head of department

Telefon +49 3677 692858

E-Mail senden




Title:ASFaLT -Automatisierte Schadstellenerkennung für unterschiedliche Fahrbahnbeläge mittels Deep Learning Techniken
Duration:01.10.2018 - 31.03.2021
Funding:Österreichische Forschungsförderungsgesellschaft mbH (FFG)
Project Partner:

AIT Austrian Institute of Technology GmbH

Project Manager TUI:Prof. Dr. Horst-Michael Groß
Staff:Dipl.-Inf. Markus Eisenbach, Dipl.-Inf. Ronny Stricker
Preceding Projects:ASINVOS


The road network in Germany, Austria, and Switzerland is aging and needs frequent condition acquisitions and assessments for timely actions for maintenance. Therefore, a frequent, network-wide analysis of the road pavement condition is required. In addition to the use for maintenance, the condition assessment is required for acceptance of construction work in large-scale projects. The image-based acquisition by measurement vehicles during regular traffic is already automated to a high degree. However, so far, the visual assessment is done manually by human experts, which screen all the recorded images and evaluate them. This process is time-consuming, exhausting, and thus, error-prone. Therefore, the objectiveness of the assessment is limited. Thus, automating the assessment is desirable and an ongoing object of research. Evaluations show that classical machine learning approaches are of limited suitability for detecting distress automatically. Additionally, they cannot detect areal damages. However, the research project ASINVOS showed, that Convolutional Neural Networks achieve promising recognition results. The detection performance is at a high level already but in order to get a practical solution.

Aims and intended results

The following extensions will be implemented as part of the project ASFaLT:

Generalization: In order to improve the generalization capabilities, we will extract a broader data basis to train the neural networks. It shall include different measurement systems and road surfaces that are typical for Germany, Austria, and Switzerland regarding condition acquisition. Machine learning approaches will help to identify which data are crucial to improve the generalization capabilities.
Quality measure: To fully automate the process, a basic necessity is to estimate the certainty of the classifier’s decisions, which is incorporated by none of the state-of-the-art systems yet. The quality measure will incorporate the heteroscedastic as well as the epistemic uncertainty. Therefore, it will be possible to evaluate if a fully automated assessment is feasible for given data.
Road damage trends: We will realize prototypical approaches to assess temporal changes of individual damages based on multiple acquisitions. In contrast to the current macroscopic point of view, we will focus on the trends of individual damage spots.
The automation will enable a consistent, objective, fast, and cost-efficient assessment. This will speed up assessments and hence, improve the maintenance schedule.