基于主干增强和特征重排的反无人机目标跟踪

Anti-UAV object tracking with enhanced backbone and feature rearrangement

视频图像中面向无人机的目标跟踪是反无人机任务中的重要一环。无人机低空飞行背景复杂,同时在视频图像中目标像素占比较小,都给目标跟踪增加了难度。针对以上问题,以SiamRPN++为基础,提出了一种引入改进的主干网络和特征重排的孪生神经网络目标跟踪算法(SiamAU)。首先,在主干网络中加入ECA-Net注意力机制网络,同时对激活函数进行改进,以提升复杂背景下的特征表征能力;然后,对主干网络输出的浅层特征进行浅层降维并与后三层深层特征进行融合,得到更适合无人机等小目标跟踪的改进深度融合特征。在DUT Anti-UAV数据集上,SiamAU算法的成功率和精确率达到了60.5%和88.1%,相比基准算法提升了5.6%和8.1%。在两个公开数据集上的测试结果表明,在反无人机场景中SiamAU算法的跟踪表现优于目前主流的算法。

Object tracking for the unmanned aerial vehicle(UAV)in videos is an important part of the Anti-UAV task.The complex background during low-altitude flight and the small imaging size are two difficulties for UAV object tracking.A Siamese neural network object tracking algorithm(SiamAU)is proposed,which is based on SiamRPN++in combination with an improved backbone and a feature rearrangement technique.Firstly,ECA-Net attention module is integrated into the backbone network,while the activation function is improved to enhance the representation ability of convolution features in complex background.Then,channel number of the last three convolution features is rearranged in order to make full use of low-level features that are conducive for small object tracking.The rearranged feathers are further fused to obtain the improved feature map.Finally,On the DUT Anti-UAV dataset,SiamAU algorithm achieves success and precession scores of 60.5%and 88.1%,an improvement of 5.6%and 8.1%in comparison with the baseline algorithm.Extensive experimental results on two public datasets validate that the proposed SiamAU achieves better UAV tracking performance and outperforms previous methods,especially in small object and complex background scenarios.