基于注意力机制改进孪生网络的无人机跟踪算法

UAV Tracking Algorithm Based on Attention Mechanism Improved Siamese Network

为了应对非合作无人机带来的安全隐患问题,对其实现有效监管,提出了一种基于改进孪生网络的无人机目标跟踪算法。首先,对全卷积孪生网络的结构进行改进,在网络结构中添加了Ghost卷积和集成卷积注意力模型;然后,利用构建的无人机数据集对改进网络算法与3种传统目标跟踪算法SiamFC、DeepSORT和FlowTrack进行训练和验证;最后,将4种算法进行对比分析,结果表明:该算法精确度达到91.4%,成功率69.6%,性能优于其他3种算法,能够有效跟踪无人机目标。

To cope with the problem of safety hazards caused by non-cooperative unmanned aerial ve-hicles(UAVs)and realize adequate supervision,a UAV target tracking algorithm based on improved siamese network is proposed.Firstly,the structure of the full-convolution siamese network(Siam-FC)is improved.Ghost convolution and integrated convolution block attention models are added to the network structure.Then,using the constructed UAV dataset,the improved network algorithm and three traditional target tracking algorithms named SiamFC,DeepSORT and FlowTrack are trained and verified.Finally,the four algorithms are compared and analyzed.The results show that the algorithm has an accuracy of 91.4%,a success rate of 69.6%,and better performance than the other three algorithms.Therefore,it can effectively track UAV targets.