Automatic Pavement Distress Detection Using a Semantics-Aided Deep Learning Method

Road maintenance plays an important role in traffic infrastructure management. Current pavement distress detection methods mostly rely on manually processing the data collected by professional instruments, which is time-consuming and expensive. In this paper, a Semantics-aided Pavement Distress Detection (SPDD) model is proposed to automatically and accurately detect the damage regions solely using the driving recorder image data. Specifically, a semantic segmentation model trained on the RTK data set is first adopted to recognize the road regions. Then, an object detection model, YOLOv5, is used to detect the damaged regions. Importantly, a novel loss function is proposed to constrain the predictions with the aid of the semantic segmentation results, encouraging the predicted boxes to be within the road regions. Detailed experiments on benchmarks show that the proposed method achieves performance competitive to those of state-of-the-art methods. Extended experiments on real-world data further demonstrate the effectiveness and practicality of the proposed method.