Patch detection for pavement assessment

Pavement management systems rely on comprehensive up-to-date road condition data to provide effective decision support for short, medium and long term maintenance scheduling. However, the cost per mile of the existing condition data collection methods allows only for periodical surveys. This leads to long gaps between inspections and a focus on major roads over rural ones. Therefore, pavement condition monitoring systems that provide inexpensive frequent updates on the road condition are necessary. Such systems would require robust and automatic defect detection methods using low-cost sensors. In this paper, one such method is proposed for detecting road patches from video data acquired by the car's parking camera. A patch is initially detected based on its visual characteristics, which are: 1) it consists of a closed contour and 2) its texture is the same with the surrounding intact pavement. The patch is then passed to a kernel tracker in order to trace it in subsequent video frames. This way redetection is avoided and each patch is reported only once. The method was implemented in a C# prototype and tested with video data consisting of approximately 4000 frames collected from roads in Cambridge, UK. The results show that the suggested method has 84% precision and 96% recall.