Automatic Pavement Crack Detection Based on YOLOv5-AH

摘要

Object detection on pavement cracks plays a significant role in the overall condition assessment of the pavement. However, the automatic crack detection remains a tough work due to the complex background and the scale change. In order to solve the related issues, a deep learning model named YOLOv5-AH is proposed on the basis of the YOLOv5. We integrate convolutional block attention module (CBAM) to extract attention area, which can emphasize the crack features and suppress the background information. Moreover, we optimize the network architecture by adding one more prediction head to detect different-scale cracks. To further improved our proposed YOLOv5-AH, a series of advanced strategies are integrated into the model training including data augmentation. A pavement crack dataset in different enviroments is built by a mobile robot. The precision and recall of our crack detection method reach 0.924 and 0.988 respectively. The prediction speed of the model can achieve real-time effect (72FPS).