遥感图像具有尺度多样、密集排列和目标尺寸小等特点，针对遥感图像中背景噪声多，车辆目标较小难以获取的问题，提出一种基于改善特征融合方法的车辆目标检测算法Atiny-YOLO。首先，在YOLOv5的Neck层中引入针对小目标的额外检测层，从而生成更大规模的特征图，有效识别小物体的细节特征；其次，向C3模块中添加split操作以复用图像特征信息，进一步优化Swin Transformer模块提高有效信息的使用率；最后，通过改善特征融合通道，提升检测精度的同时减少模型参数。在无人机视角数据集（aerial universal autonomous inspection and recognition, AU-AIR）数据集上验证Atiny-YOLO模型的有效性。实验结果表明：Atiny-YOLO算法相较于基线算法的平均检测精度提高了约2.9%。达到95.5%，检测速度达到234帧／s。这些结果验证了Atiny-YOLO算法在满足实时性的同时，模型检测精度大幅提升。

Remote sensing images are characterized by diverse scales, dense arrangement and small target sizes, etc. Aiming at the problem that there is much background noise in remote sensing images and vehicle targets are small and difficult to be acquired. A vehicle target detection algorithm based on improved feature fusion method, Atiny-YOLO was proposed. Firstly, an additional detection layer for small targets was introduced into the Neck layer of YOLOv5 so as to generated a small target detection algorithm for drone remote sensing images. Neck layer to introduce an additional detection layer for small targets, so as to generated a larger-scale feature map and effectively identified the detailed features of small objects. Secondly, a split operation was added to the C3 module to reuse the image feature information, and the Swin Transformer module was further optimized to improve the usage rate of the effective information. Lastly, by improving the feature fusion channel, the detection accuracy was improved while the model parameters were reducing the model parameters. The Atiny-YOLO algorithm was tested on the AU-AIR (aerial universal autonomous inspection and recognition) dataset. The experimental results show that the average detection accuracy of the Atiny-YOLO algorithm compared to the baseline algorithm is improved by about 2.9%. It reaches 95.5% and the detection speed reaches 234 frames/s. These results verify that the Atiny-YOLO algorithm meets the real-time performance while the model detection accuracy is greatly improved.