航空学报 > 2025, Vol. 46 Issue (16): 331619-331619   doi: 10.7527/S1000-6893.2025.31619

低空无人机实时目标检测算法

杨永刚, 姜文韬, 高志云()   

  1. 中国民航大学 交通科学与工程学院,天津 300300
  • 收稿日期:2024-12-06 修回日期:2024-12-27 接受日期:2025-03-05 出版日期:2025-03-19 发布日期:2025-03-19
  • 通讯作者: 高志云 E-mail:zygao@cauc.edu.cn
  • 基金资助:
    国家自然科学基金(62403471);中央高校基本科研业务费(3122023QD18);天津市城市空中交通系统技术与装备重点实验室(TJKL-UAM-202402)

Real-time target detection algorithm for low altitude UAVs

Yonggang YANG, Wentao JIANG, Zhiyun GAO()   

  1. School of Transportation Science and Engineering,Civil Aviation University of China,Tianjin 300300,China
  • Received:2024-12-06 Revised:2024-12-27 Accepted:2025-03-05 Online:2025-03-19 Published:2025-03-19
  • Contact: Zhiyun GAO E-mail:zygao@cauc.edu.cn
  • Supported by:
    National Natural Science Foundation of China(62403471);The Fundamental Research Funds for the Central Universities(3122023QD18);Key Laboratory of Technology and Equipment of Tianjin Urban Air Transportation System(TJKL-UAM-202402)

摘要:

针对低空无人机视角下的目标存在相互遮挡、像素小和复杂背景的问题,提出一种用于低空无人机平台的小目标检测算法HPRS-YOLO。在主干网络采用一种新的多尺度空间金字塔(SPMCC),抛弃基于最大池化的下采样形式,利用膨胀卷积动态调整网络的感受野,更有效地绘制检测对象的上下文信息;融合2种Metaformer模型改进C3K2模块,增强小目标结构和纹理特征信息,减少参数量,保持运算开销在较小水平;Dysample优化上采样算子,抑制偏移重叠和边界点值混乱,提高目标与背景的对比度;引入浅层细节处理模块(SDFM)重新设计颈部网络尾端,实现首尾跨尺度特征校准,强调对低层特征图的关注度,补偿小目标特征的缺失以及维护遮挡目标剩余空间信息的完整性。对数据集VisDrone2019进行消融实验和对比实验,相较于基线算法,mAP0.5和mAP0.5∶0.95分别提升5%和3%,对公开数据集DOTA进行泛化实验,mAP0.5提升2.0%,证明了所提算法具有良好的鲁棒性,最后将模型部署到嵌入式设备NVIDIA Jetson AGX Orin上进行验证,FPS达到60,表明HPRS-YOLO通过优化算法设计可以在保持高准确率的同时,确保实时检测的能力。

关键词: 低空无人机, 小目标检测, 多尺度, 跨尺度特征校准, YOLOv11n, Jetson AGX Orin

Abstract:

To address the challenges of mutual occlusion, tiny pixels, and complex backgrounds in low-altitude UAV-based object detection, this paper proposes HPRS-YOLO, a small target detection algorithm optimized for UAV platforms. The backbone network incorporates a novel Spatial Pyramid Multi-scale Common Convolution (SPMCC), which replaces max-pooling-based downsampling with dilated convolution to dynamically adjust the receptive field, thereby enhancing contextual feature extraction. The improved C3K2 module integrates two Metaformer architectures to reinforce structural and textural features of small targets while reducing parameters and maintaining low computational overhead. Additionally, a dynamic upsampling operator, Dysample is introduced to suppress offset overlaps and boundary pixel value confusion, thereby improving target-background contrast. The neck network is redesigned with a Shallow Detail Focus Module (SDFM) to achieve cross-scale feature calibration between terminal layers, emphasizing low-level feature maps to compensate for missing small-target characteristics and preserve spatial integrity of occluded objects. On the dataset VisDrone2019, ablation and comparison experiments are conducted. The results show that mAP0.5 and mAP0.5∶0.95 are improved by 5% and 3%, respectively, when compared to the baseline method. Generalization experiments are conducted on the public datasets DOTA, and mAP0.5 is improved by 2.0%, demonstrating good robustness. Finally, the model is deploying the model on an embedded NVIDIA Jetson AGX Orin device achieves an FPS of 60, demonstrating that HPRS-YOLO guarantees real-time detection capability by optimizing the algorithm design while keeping high accuracy.

Key words: low-altitude UAV, small target detection, multi-scale, cross-scale feature calibration, YOLOv11n, Jetson AGX Orin

中图分类号: