注意力聚集无锚框的孪生网络无人机跟踪算法

Attention aggregation siamese network with anchor free scheme forUAV object tracking

针对无人机目标跟踪过程中经常出现角度变化、形变、相似物体干扰等问题,提出轻量级注意力聚集无锚框的孪生网络无人机实时目标跟踪算法.考虑到无人机高空视角跟踪目标较小,在特征模板两分支中引入高效通道注意力机制,能够有效获取目标的语义信息和细节信息.在融合两层响应的基础上,引入空间注意力机制,能够有效地聚合注意力特征,同时扩大模型的视野范围.引入无锚框机制,针对每个像素进行分类和预测回归目标框,减少了模型复杂度,大大降低了计算量.在UAV123@10fps、UAV20L和DTB70等无人机跟踪数据集上与多个当前比较流行的算法进行对比实验,结果表明,所提算法在3个无人机数据集上的平均跟踪速度达到155.2帧/s,在多种复杂环境下,均能实现对目标的有效跟踪.

A real-time UAV object tracker based on lightweight and attentional aggregation siamese network with anchor free scheme was proposed,aiming at the problems of viewpoint change,deformation and similar objects around in UAV tracking tasks.Considering the small number of object pixels in the view of UAV high-altitude platform,an efficient channel attention scheme was introduced to the two branches of template.Then semantics information and detail information can be effectively extracted.A spatial attention scheme was constructed to effectively aggregate attention and enlarge the visual field range after fusing the response of two layer.An anchor free mechanism was built to directly classify and predict the object box on each pixel,which can simplify the complexity of model and reduce the calculation cost.The proposed method was conducted on three public UAV data sets such as UAV123@10fps、UAV20L and DTB70,and compared with other state-of-the-art tracking algorithms.The experimental results show that the proposed method can track the target effectively in many challenging scenes with an average speed of 155.2 frame per second on three UAV benchmarks.