Quality Assessment of Flexible Pavement Automated Cracking Data in Louisiana

Abstract

Pavement distress survey is the most critical and integral part of pavement management system. Transportation agencies all over the world depends on the accurate pavement distress survey to take major decisions like pavement rehabilitation or reconstruction. Pavement management at the network level requires accurate pavement distress survey to decide the allocation of budget for pavement maintenance, rehabilitation and reconstruction. This study started collected network-level cracking distress data through manual measurements on high-resolution pavement images and the results were used as the ground truth measurements to evaluate the corresponding automated distress data from the Pavement Management System database. Results showed that the automated cracking measurements based on 0.1-mile long section tend to largely over-estimate all types of cracking (alligator, longitudinal and transverse) at different severity levels, especially the amount of cracking at a moderate severity level. On the other hand, when the automated cracking measurements were reported based on 50-ft long section, the overall estimation errors without differentiating the cracking severity levels were significantly reduced due to a smaller standard deviation of the measurement errors within a shorter survey section. Furthermore, automated measurements at project-level and network-level were statistically compared with the ‘ground truth’ manual measurements. Statistical t-test analysis on the mean measurement error and equality of variance analysis were conducted to qualitatively evaluate the accuracy and precision of the data. Statistical analyses indicated that, in general, automated measurements were significantly different than the manual measurements. To improve the accuracy of the automated cracking measurements, an artificial neural network (ANN) model was developed using the manual and automated cracking measurements from 23 asphalt pavement sections in Louisiana. The ANN model aimed at adjusting the less accurate automated cracking measurements towards more accurate manual cracking measurements. Evaluation of the developed ANN model indicated that the predicted cracking measurements correlated well with the manual measurements and using the predicted results can produce a similar cracking indices as those from the manual measurements. Moreover, Louisiana Transportation Research Center (LTRC) has its’ own digital highway data collection vehicle which can collect high resolution pavement images at highway speeds. However, the supporting software cannot provide automated cracking measurements. In this study, an automated pavement crack survey procedure was developed to utilize the high resolution pavement images collected by LTRC. ‘Structured Forests Algorithm’ was used for crack detection. Crack classification and quantification algorithms were developed and integrated with the crack detection algorithm using MATLAB.