Structural Defect Detection for Urban Road Pavement Using 3D Ground Penetrating Radar Based on Deep Learning

The prevention of road collapse accidents caused by hidden structural defects becomes an urgent problem for road safety. Three-dimensional ground penetrating radar (3D GPR) is an advanced non-destructive detection approach to effectively detecting road hidden defects. However, the GPR images are difficult to interpret, and the manual interpretation speed is slow. To realize the automatic location and recognition technology of road internal diseases and improve detection efficiency, a large number of measured 3D GPR road data are used to establish a B-scan image disease database in this study. Data preprocessing, image capturing, defects marking, and data cleaning are performed in this database. Deep learning convolution neural network models were built based on one-stage methods (YOLOv3 and YOLOv4) and a two-stage method (Faster R-CNN). Through comparing and analyzing their recognition effect and performance differences. The frames per second (FPS) of YOLOv3 and YOLOv4 are much larger than that of Faster R-CNN. Generally, the YOLOv4 has the best performance among all the models, and the prediction accuracy of four features from high to low is well, crack, concave, cavity.