【摘要】 随着无人机技术的快速发展，航拍图像的车辆检测逐渐成为计算机视觉研究的热点。因为航拍位置较高，导致航拍图像中的小目标较多。针对小目标检测难问题，在YOLOv3模型的基础上进行优化，提出一种航拍视觉下的车辆检测模型(Vehicle-YOLO),通过采用原图下采样4倍、8倍和16倍特征，引入SPP模块和将原YOLOv3中的Convset模块改为残差连接的形式等优化机制来提升模型的目标检测性能。分别利用YOLOv3-tiny、YOLOv3和所提出的Vehicle-YOLO模型对航拍图像车辆进行检测，从模型检测精度上分析，YOLOv3-tiny、YOLOv3和Vehicle-YOLO检测精度分别为68.38%、75.45%和88.74%。所提出的Vehicle-YOLO模型较YOLOv3的精度高了13.29%,较YOLOv3-tiny的精度高了20.36%。在单张2080Ti的GPU平台上，YOLOv3-tiny、YOLOv3和Vehicle-YOLO检测速度分别为113 FPS、42 FPS和39 FPS。可见，Vehicle-YOLO在精度和速度上取得了较好的平衡。更多还原

【Abstract】 With the rapid development of unmanned aerial vehicle technology, vehicle detection of aerial images has gradually become a hot research direction of computer vision. Because of the high position of the aerial photograph, there are many small objects in the aerial image. In order to solve the problem of small object detection difficulty, this paper optimized the YOLOv3 model, and put forward a vehicle detection model(Vehicle-YOLO) under aerial photography vision, which used 4, 8, and 16 times of the original sample characteristics. The SPP module and the Convset module in YOLOv3 were introduced to improve the object detection performance of the model. YOLOv3-tiny, YOLOv3 and Vehicle-YOLO models were used to detect vehicles in aerial images. From the accuracy analysis of model detection, the detection accuracies of YOLOv3-tiny, YOLOv3 and Vehicle-YOLO are 68.38%, 75.45% and 88.74%, respectively. The accuracy of Vehicle-YOLO model is 13.29% higher than that of YOLOv3 and 20.36% higher than that of YOLOv3-tiny. On a single 2080Ti GPU platform, the detection speeds of YOLOv3-tiny, YOLOv3 and Vehicle-YOLO are 113FPS, 42FPS and 39 FPS, respectively. It can be seen that Vehicle-YOLO achieves a good balance in accuracy and speed.更多还原