基于双注意力混洗的无人机航拍目标跟踪算法

针对无人机（UAV）跟踪过程中目标经常出现尺寸小、尺度变化大和相似物干扰等问题，提出了一种基于双注意力混洗的多尺度无人机实时跟踪算法。考虑到无人机视角下目标像素点少，构建了双采样融合的深层网络，既提供了语义信息丰富的深度特征,又保留了目标的细节信息；设计了双注意力混洗模块，通道注意力和空间注意力同时分组筛选提取到的特征信息，混洗不同通道间的信息，加强信息交流，提高了算法辨别能力；为利用不同层的特征信息，加入多个区域建议网络完成目标的分类和回归，并针对无人机的目标特点，将结果进行加权融合。实验结果表明：所提算法在数据集上的成功率和准确率分别为60.3%和79.3%，速度为37.5帧/s。所提算法的辨别能力和多尺度适应能力明显增强，能有效应对无人机跟踪中常见的挑战。

Aerial object tracking algorithm for UAVs based on dual-attention shuffling

A multi-scale real-time tracking algorithm for unmanned aerial vehicle (UAV) based on dual-attention shuffling is proposed to solve the problems of small size, large scale variation and similar object interference which often occur during UAV object tracking. First, considering the small number of target pixels in the UAV view, a deep network with double sampling integration is constructed, which provides semantic information-rich depth features and preserves the target’s detailed information. Next, a dual-attention shuffling module is designed. Channel attention and spatial attention are simultaneously grouped to filter the extracted feature information, and then the information between different channels is shuffled to enhance information exchange and improve the discriminative ability of the algorithm. Finally, to utilize the feature information of different layers, multiple region proposal networks are added to complete the target classification and regression, and the results are weighted and fused for the UAV target characteristics. Results show that the success and precision rates of the algorithm are 60.3% and 79.3% on the dataset, respectively, with 37.5 frame/s. The algorithm discrimination ability and multi-scale adaptation are significantly enhanced, which can effectively deal with the common challenges in UAV tracking.