Image Recognition of Pavement Cracks in Autonomous Driving Scenarios Based on Deep Learning

Pavement cracking represents the predominant form of defect encountered within autonomous driving environments. With the advancement of deep learning technologies, an expanding array of methods has been deployed for the extraction of crack information from pavement imagery. Addressing the challenges of insufficient accuracy and the absence of real-time capabilities, attributable to the incomplete extraction of crack features by existing deep learning-based road crack recognition methodologies, this paper designs a novel pavement crack detection approach that utilizes the enhanced YOLOv5s model. In the encoding and decoding sections, the Efficient Channel Attention (ECA) module and residual connections were incorporated to enhance feature extraction capabilities. This approach not only extracts local feature information of road surface cracks but also ensures the model’s lightweight architecture. The network, once trained, extracts feature from various types of crack images to ultimately classify and detect cracks in the pavement. The experimental findings revealed that the model is capable of efficiently recognizing several distinct types of road cracks, achieving a recognition mAP of $\mathbf{9 2. 3 \%}$ and an F 1 score of $\mathbf{9 1. 3 \%}$. In scenarios with varying lighting and viewing angles, the model demonstrates rapid detection speed, high accuracy, and strong robustness, fulfilling the reliability needs for crack detection and identification in autonomous driving contexts.