Aiming at the problem that the target size of the current UAV aerial images is small and the image background is complex, which leads to low detection accuracy of existing UAV target detection algorithms, this article proposes an improved YOLOv8s UAV target detection algorithm. First, deformable convolution is used to replace standard convolution to enhance the network's feature extraction ability for irregularly shaped targets; Then the separable large-core attention mechanism LSKA is used to improve the SPPF module to improve the problem of low detection accuracy due to large differences in target scales. The weighted bi-directional feature pyramid network Bi-FPN is combined at the neck of the network to achieve multi-scale feature fusion and improve the network's missed detection and false detection of small targets. At the head of the network, the dynamic detection head DyHead is used to replace the original detection head to enhance the detection ability of occluded objects and small targets. Finally, in order to solve the problem that a large number of low-quality samples in the dataset have a negative impact on the training process, the Wise-IOU loss function was used to improve the model convergence speed and detection accuracy. Experimental results show that the improved method achieved 41.7% mAP on the VisDrone2019 dataset. Compared with the original YOLOv8s algorithm, mAP50 increased by 3.0%, mAP50:95 increased by 1.9%, the number of parameters decreased by 17.5%, and the amount of calculation dropped by 12.63%. It achieves both model lightweight and detection accuracy.