In view of the low identification accuracy of the crack detection technology of asphalt pavement under the current conditions of complex pavement, an asphalt pavement crack detection algorithm based on improved YOLOv5s was proposed by building the dataset for asphalt pavement cracks. The first step was to make the following improvements to the original YOLOv5s model according to the characteristics of the asphalt pavement crack dataset: obtain the initial anchor box matched with the characteristics of the asphalt pavement cracks by re-clustering the anchor box of the crack dataset using the K-means + + algorithm; add the convolutional block attention module(CBAM)to the Prediction part of the model to improve the model's ability to detect small cracks; take the CIoU_Loss function as the model regression loss function to improve the anchor box location accuracy. The second step was to perform an ablation experiment on the improved YOLOv5s model, Which would prove that each improvement scheme could improve detection ability without conflict. The final step was to compare the improved YOLOv5s model with various classic target detection models in the dataset of this paper: the CFD dataset, the Crack500 dataset and the Crack200 dataset. The results show that the detection of the improved YOLOv5s model on each dataset was better than other target detection models. The mAP@0.5 and mAP@[0.5:0.95] of this model on the dataset were 90.58% and 56.08%, which were much higher than other target detection models. These findings indicate that the improved YOLOv5s model had better detection under complex pavement conditions and could provide a theoretical basis for the automatic detection of asphalt pavement cracks.