Road defect detection is a crucial aspect of road maintenance projects, but traditional manual methods are time-consuming, labor-intensive, and lack accuracy. Leveraging deep learning frameworks for object detection offers a promising solution to these challenges. However, the complexity of backgrounds, low resolution, and similarity of cracks make detecting road cracks with high accuracy challenging. To address these issues, a novel road crack detection algorithm, termed Road Defect Detection YOLOv5 (RDD-YOLOv5), was proposed. Firstly, a model was proposed to integrate the transformer structure and explicit vision center to capture the long-distance dependency and aggregate key characteristics. Additionally, the Sigmoid-weighted linear activations in YOLOv5 were replaced with the Gaussian Error Linear Units to enhance the model's nonlinear fitting capability. To evaluate the algorithm's performance, a UAV flight platform was constructed, and experimental freebies were provided to boost inspection efficiency. The experimental results demonstrate the effectiveness of RDD-YOLOv5, achieving a mean average precision of 91.48%, surpassing the original YOLOv5 by 2.5%. The proposed model proves its ability to accurately identify road cracks, even under challenging and complex traffic backgrounds. This advancement in road crack detection technology has significant implications for improving road maintenance and safety.
Keywords: GELU; YOLOv5; deep learning; road crack detection; transformer.