Road crack detection of drone-captured images based on TPH-YOLOv5

Pavement cracks are common indicators of road distress. Road maintenance agencies must monitor conditions and detect cracks promptly to reduce costs, ensure structural reliability, and enhance safety and comfort for drivers. Road crack detection is typically performed using road inspection vehicles with human drivers, but this approach is inefficient for large-scale road networks, consuming significant manpower and time. Utilising unmanned aerial vehicles (UAVs) for crack detection can save time and manpower without disrupting traffic flow. However, images of cracks captured by UAVs often have complex backgrounds and unevenly distributed objects, which pose challenges for existing models in achieving the required accuracy in real road environments. The Transformer Prediction Heads-YOLOv5 (TPH-YOLOv5) model is employed to address these issues in this paper. It improves the YOLOv5 model by adding a prediction head for better crack detection across different sizes. The original detection head is upgraded to TPH integrated with a convolutional neural network for multi-scale feature fusion. This design enhances the model's detection capability for cracks of varying sizes and improves robustness. Additionally, the model incorporates the Convolutional Block Attention Model (CBAM) for adaptive channel and spatial attention, enhancing performance in identifying road cracks. By employing data augmentation, multi-scale testing, model integration, and additional classifiers, the model achieves superior performance. Experimental results on the RDD2022 China_Drone dataset demonstrate improved accuracy, recall, and mAP metrics compared with YOLOv5, meeting engineering application requirements.