Lightweight Yet Precized: A Redesigned YOLO for High-Accuracy Road Crack Detection on Vehicle-Mounted Devices

Road crack detection plays a critical role in ensuring traffic safety and enabling timely maintenance. However, real-time detection algorithms deployed on vehicle-mounted platforms often struggle with several challenges. These include poor sensitivity to small cracks, incomplete feature extraction, and difficulty balancing detection accuracy with model efficiency. This study presents LP-YOLOv8, an enhanced object detection algorithm based on an improved YOLOv8n architecture. The proposed method is designed to offer a lightweight yet precise solution for road crack detection. To reduce model complexity while maintaining high accuracy, we introduce the C2f-faster module, which optimizes the backbone network by minimizing parameter redundancy and suppressing background noise. We further propose a Lightweight Shared Detail-Enhanced Convolution Detection Head (LSDECD) to enhance multi-scale feature perception while minimizing parameter count. To improve the detection of small cracks, we redesign the neck using a Focal Diffusion Pyramid Network (FDPN). FDPN strengthens cross-scale feature fusion by regulating deep-to-shallow feature interactions through learnable attention gating. Moreover, we propose an improved Inner-WIoU loss function that focuses on small objects and samples with general quality labels. The LP-YOLOv8 method was evaluated on the RDD2022 vehicle-mounted image dataset through extensive comparative and ablation studies. Experimental results demonstrate a 3.41% improvement in mAP@0.5 over YOLOv8n, while reducing the number of parameters by 55.6% and computational costs by 36.7%. These findings confirm that LP-YOLOv8 provides an efficient and accurate solution for real-time road crack detection in vehicle-mounted imaging systems.