Deep Learning-Based Detection of Tie Bars in Concrete Pavement Using Ground Penetrating Radar

Xuetang Xiong,Yiqiu Tan

DOI: https://doi.org/10.1080/10298436.2022.2155648

2022-01-01

Abstract:Detection of the tie bars in concrete pavement has been a challenging task. To address the purpose, ground penetrating radar (GPR) was used to acquire a large amount of image data along the longitudinal construction joints of plain concrete pavement in the field. The GPR image data was filtered to construct the dataset, containing 2185 tie bar reflected waves in 670 GPR images. Then, the YOLO series models, as the deep learning algorithms applied in inspecting the tie bars from GPR images, were well trained with the GPR training and validation sets. The comprehensive detection accuracy of the YOLOv4 model outperforms the YOLOv3, YOLOv3-tiny, and YOLOv4-tiny models in the test set. The mAP@0.5 value of the YOLOv4 model can reach 99.74%. All the signatures of tie bars in the testing GPR images, no matter whether they are incomplete, compressed, blurry with missing signal, or strong background noise, can be correctly and completely anchored using the bounding box based on the YOLOv4 model. Meanwhile, the detection speed of the YOLOv4 model for GPR data video is 50.8 frames per second. Therefore, the YOLOv4 model is reliable for automatically detecting the tie bars from GPR data in real-time.