This research project aimed to identify a cost-effective method for collecting distress data on secondary roadways and to develop efficient deep learning models for classifying and quantifying pavement cracks on both primary and secondary roadways in North Carolina. The project utilized high-resolution images provided by the NCDOT as well as images captured using GoPro cameras.
The findings revealed that GoPro cameras, when mounted on the rear of a vehicle, driven at speeds below 20 mph, and used in fair weather conditions, offer a low-cost solution for data collection. For high-accuracy image annotation, Fiji (ImageJ) with its Trainable Weka Segmentation (TWS) plugin proved to be highly effective.
Recommended Setup of a GoPro Camera
For crack classification on primary roadways, both ResNet and Vision Transformer models demonstrated exceptional performance, achieving a 97% accuracy in identifying crack types (longitudinal, transverse, and alligator cracks). Similarly, for secondary roadways, ResNet and Vision Transformer models are recommended, achieving an 85% accuracy using images extracted from GoPro videos. In terms of crack segmentation, U-Net and DeepSegmentor models are recommended for primary roadways, with Dice Coefficients of 97%. For secondary roadways, the DeepSegmentor model is recommended due to its superior performance in handling complex crack patterns. For crack quantification, the use of pixel-level segmentation with real-world calibration ensured precise measurements of crack length, width, and area.
Once implemented, the outcomes of this research have the potential to significantly enhance the current NCDOT Pavement Management System (PMS). This will enable more frequent and cost-effective monitoring of roadway conditions, improve the accuracy of pavement performance predictions, and assist engineers in maintaining roadways with enhanced performance, extended lifespan, and reduced maintenance requirements.
