University of Alcalá
|Proposed start date:||2008-09-26||Proposed end date:||2011-09-26|
A quality road network has a great importance to the national economy and preserving its condition becomes a priority for Governments and institutions. Poorly maintained road networks expose users to unacceptable health and safety risks increasing the number of accidents as well as pollution and noise levels. A preventive maintenance policy has proved to be a cost-effective strategy, reducing up to five times the rehabilitation costs once the road is fully deteriorated.
When evaluating the road quality, cracking has arisen as the best indicator of the need of preventive maintenance treatments. However, evaluating the road quality is a tedious, laborious and slow work, that requires several operators to collect the information and manually analyze it with the well-known problems with inconsistency and repeatability of the results.
In the last years, great effort has been put toward the development of a fully automated distress detection system to give a quantitative measure of the quality of the road surface and assist prioritizing and planning the maintenance of the road network.In this thesis a complete vision-based road distress detection framework, including the phases needed to properly deal with a fully automatic road distress assessment is presented.
One of the main challenges for a fully automatic system is to cope with the different classes of road surfaces present in a road network and tune the system parameters to detect cracking on very different textures. A linear SVM-based multi-class classifier has been trained and tested to distinguish between 10 different types of road surfaces that appear in the Spanish roads. After an exhaustive study a feature vector including Gray Level Co-occurrence Matrix and Local Binary Pattern has been used to classify the road surface achieving high detection rates. In a further step, non-crack features such as joints, road markings and sealed cracks are detected and not considered as cracking. Once the non-crack features are removed, cracking is extracted using a complete crack detector module based on Multiple Directional Non-Minimum Suppression seed extraction and minimum cost path growing. Then, cracks are classified into different types according to their topologies and severities. The crack detection process involves the use of several finely tuned parameters. These parameters are tuned depending on the output provided by the classifier.The overall system performance as well as the influence on the final results of the non-crack detector and the road surface classifier module, have been tested with an extensive dataset of manually labeled road images. Results for an exploitation scenario with a survey of more than three thousand kilometers of the Spanish road network are presented and discussed.