Monocular Road Damage Size Estimation using Publicly Available Datasets and Dashcam Imagery

Abstract

Among the challenges of maintaining a safe and efficient transportation system, Departments of Transportation (DOT) must assess the quality of hundreds-of-thousands of miles of roadway every year and prioritize limited resources to address issues that affect safety and reliability. In particular, road damage in the form of 3D analysis of cracks and potholes is difficult to catalog and require significant human resources to survey. However, a new and growing remote-sensing network comprised of low-cost consumer dashcams presents an opportunity to dramatically lower the cost and effort required to perform road damage assessments. This paper provides methods to approach this problem and details a number of public datasets and models that can be used to tackle it. The central contribution here is a set of several practical software pipelines designed to accomplish this task in an automated fashion. An emphasis on deep learning methods is presented that enables organizations to improve or tailor the results according to their specific requirements and the availability of labeled data. Suggestions for possible directions for future work and improvements at each stage of the pipeline are also presented.