结合仿真迁移学习和自适应融合的无人机小目标检测

Combination of Simulation-based Transfer Learning and Adaptive Fusion for UAV Small Object Detection

无人机小目标的精确检测在公共安全和无人机防御系统中起着至关重要的作用.被广泛应用于通用目标检测任务的深度学习技术,在无人机小目标检测任务上的效果往往受限于稀缺的相关数据资源以及较小的目标尺度.针对以上问题,本文提出了一种基于仿真迁移学习和自适应融合机制相结合的无人机小目标检测方法.该方法首先利用基于UnrealEngine的Air-Sim仿真平台生成丰富且高保真的无人机小目标仿真图像数据,以减轻对稀缺真实图像数据的依赖.其次,为解决仿真图像与真实图像的数据分布差异问题,应用模型参数知识迁移技术,首先在仿真数据集上YOLOv5目标检测模型进行预训练,随后利用真实数据集对模型进行微调训练.最后,为进一步适应小目标检测场景,提出了一种基于YOLOv5的改进神经网络模型AF-YOLO,该网络引入了自适应融合机制.实验结果表明,基于仿真的迁移学习方法效果优于基准方法,使无人机目标检测的性能提升2.7%;引入自适应融合机制的方法,使性能提升6.2%;最终,基于仿真迁移学习和自适应融合机制相结合的方法与基准方法相比,性能提升7.1%.

Precisely detection of Unmanned Aerial Vehicles(UAVs)small objects plays a critical role in public security and UAV de-fense systems.Deep learning,which is widely adopted for general object detection tasks,shows poor performance on UAV small object detection because of the scarce relevant data resources and UAVs'small size.A novel and comprehensive approach for UAV small ob-ject detection which combines transfer learning based on simulation data and adaptive fusion is proposed to solve the problems.Firstly,the simulation platform based on AirSim and Unreal Engine 4 is used to generate plenty of UAV small object simulation images with high fidelity to reduce the dependence on real-world image data.Secondly,to close the gap between the simulation and real-world da-ta,model parameter based transfer learning is applied to pre-train a YOLO v5 model on the simulation data and fine-tune it on the real-world data.Finally,a new network model called AF-YOLO based on YOLO v5 is proposed,which introduces an adaptive fusion mechanism to improve the performance of small object detection.The experiment results show that the simulation-based transfer learn-ing outperforms the baseline with a 2.7%performance increase.Furthermore,the proposed AF-YOLO model achieves a 6.2%im-provement compared with YOLO v5.Finally,the approach for UAV small object detection with combination of transfer learning based on simulation data and adaptive fusion,brings an improvement of 7.2%compared with the baseline method.