Autonomous pavement distress detection using ground penetrating radar and region-based deep learning

Many data processing technologies have been utilized for pavement distress detection (e.g., reflection cracks, water-damage pits, and uneven settlements) using ground penetrating radar (GPR). However, the various real-world conditions have resulted in challenges of the accuracy and generalization ability of these techniques. To overcome these challenges, we proposed a deep-learning method, called faster region convolutional neural network (Faster R-ConvNet), to complete the task. The 30 Faster R-ConvNets were trained, validated, and tested using 2,557, 614, and 614 GPR images, respectively. The optimal anchor size and ratio were determined based on the validation results. The stability, superiority, real-time of the optimal Faster R-ConvNet were verified based on the test results. The results demonstrated that the optimal Faster R-ConvNet achieved 89.13% precision and 86.24% IoU. The stability of the model in different pavement structures was desirable. The comparative study indicated that the optimal Faster R-ConvNet outperformed other supervised learning methods in distress detection. Additionally, a real-time detection using optimal Faster R-ConvNet was conducted with acceptable accuracy.