Aiming at the problems of low accuracy, poor positioning ability, and large computational load in existing pavement crack detection, this study investigates a lightweight detection model based on YOLO v8n. GhostNet v2 is adopted as the backbone network to reduce the number of parameters and improve feature extraction performance. The CBAM (Convolutional Block Attention Module) is introduced into the feature fusion module to enhance attention to pavement defect features, and the bounding box regression loss function is replaced with Focal-SIoU to optimize the calculation of bounding box overlap. Experimental results show that the precision, recall, and mean average precision of the improved model are increased by 4.76%, 1.89%, and 2.77%, respectively. The model size is 33.33% of that of YOLO v8n, with an inference speed of 271.5 frames·s-1, and it takes only 3.68 seconds to detect 1000 images. The model meets real-time requirements, reduces missed and false detections of pavement cracks, and provides technical support for automatic detection on mobile pavement devices.