skip to content

Automated Patch Detection for Pavement Assessment

Stefania C. Radopoulou, PhD Candidate is focusing on automating the process of road condition monitoring. Current practice includes laborious and time-consuming manual visual surveys. Additionally, it has been proved that the subjectivity of the inspector is inevitable to affect the assessment results. Although dedicated vehicles equipped with several sensors that automatically collect data exist, their high operational and purchase cost restrict their usage to the primary road network and only once a year. This leads to long gaps between inspections and a focus on major roads over rural ones. The Department of Transport and Highways Agency, report that there are gaps in the information they have regarding road condition and that the data they own is insufficient.

One of the objectives of this research is to propose a method that is capable of detecting and tracking patches in pavement surface videos while maintaining the lowest cost possible. To do so, the use of video feed collected by parking cameras that already exist in many modern cars is proposed. The motivation lies in the idea of potentially crowdsourcing the task of monitoring to everyday road users by transforming them into ubiquitous condition reporters.

The proposed method is created based on the framework of Visual Pattern Recognition models proposed by Dr Brilakis and previous students of his (Dr German and Dr Zhu) for creating models to automate the detection of infrastructure-related objects utilizing their distinctive characteristics such as metric measurements, and/or geometric properties. The main visual characteristics used for the detection of a patch are the following: a) it consists of a closed contour, and b) the texture of a patch is similar to the pavement surrounding it. Once a patch is detected, it’s passed to a kernel tracker to track it in subsequent video frames.

The algorithm was validated with data collected from the local streets of Cambridge, UK. The experimentation results provided an accuracy of 75% for the detection, with 82% precision and 86% recall. As for the overall proposed method, including both the detection and tracking of patches, the precision is 84% and the recall is 96%.

Further details about this work can be found here.