Abstract:

The intelligent application of unmanned aerial vehicles （UAVs） is a hot topic in current research. To achieve efficient and real-time air-to-ground target detection for UAVs， this paper proposes a lightweight MBCF-YOLO model for air-to-ground target detection. First， a deep separable convolution to redesign an efficient and lightweight backbone network is introduced. Then， a small target detection branch with embedded recursive attention mechanism is added to the neck network to improve the detection accuracy of small targets and the ability of feature fusion. Finally， the focus loss function is combined with loss rank mining to improve the original loss function to overcome the data imbalance problem of air-to-ground datasets. Several control experiments and real machine applications have been conducted on this algorithm model， and the results show that compared with other current algorithms， the MBCF-YOLO algorithm achieves a better balance between accuracy and latency on the VisDrone 2021 dataset.

Key words: air to ground target detection, network lightweight, circulatory attention mechanism, small target detection, feature fusion

摘要：

无人机的智能应用是当下研究的热点，为实现高效实时的无人机对地目标检测，提出了一种应用于边缘设备的轻量级空对地目标检测算法MBCF-YOLO。引入深度可分离卷积，优化原有的骨干网络。在颈部网络中添加嵌入循环注意力机制的小目标检测分支，提高对小微目标的检测精度和特征融合的能力。将焦点损失函数与损失秩挖掘结合，改进原有的损失函数，以改善空对地数据集的数据失衡问题。对该算法模型进行了多组对照实验以及实机应用，结果表明，与当前其他算法相比，MBCF-YOLO算法在VisDrone2021数据集上的准确性和延迟之间实现了更好的平衡。

关键词: 空对地目标检测, 网络轻量化, 循环注意力机制, 小目标检测, 特征融合