Detection of Multiple Road Defects for Pavement Condition Assessment

Abstract

Pavement condition assessment is a prerequisite for pavement maintenance. State of the art practices are mainly manual, with only 0.02% of inspections using expensive specialized vehicles that automate data collection. This expensive and time-intensive task could be crowdsourced by transforming every day road users into omnipresent pavement inspectors using pre-existing car sensors. The objective is to automate the detection of pavement defects. This paper presents an application of the Semantic Texton Forests (STF) algorithm for automatically detecting patches, potholes and three types of cracks in video frames captured by a common parking camera. The prototype has been implemented in C# and validated using real data collected from local roads in the city of Cambridge, United Kingdom. Preliminary results demonstrate the successful application of STF in the context of pavements with over 70% accuracy in all of the tests performed, and over 75% precision for most of the defects.