Abstract: To enable autonomous driving systems to effectively detect and locate road potholes in complex environments, improvements have been made to the existing YOLOv5s object detection algorithm. Firstly, MobileNetV3 is employed to replace the original backbone of the model, reducing the parameter counts and achieving a more lightweight network design. Additionally, a BiFPN (bidirectional feature pyramid network) module is introduced in the neck of YOLOv5s, significantly enhancing the model’s performance in multi-scale feature fusion, information propagation, feature representation, and detection accuracy, while maintaining the lightweight nature of the architecture. Furthermore, the concept of image style transfer from generative adversarial networks is incorporated, utilizing PaddleGAN for data augmentation to increase the diversity of the dataset. Finally, experiments conducted on a custom dataset revealed that the improved YOLOv5s algorithm exhibits a 0.035 enhancement in accuracy and a 0.009 increase in mean average precision mAP50, achieving a detection speed of 54.1 frames per second. The proposed algorithm is more lightweight and enhances detection precision, providing a valuable technical reference for pothole detection in complex environments.