基于多层特征自适应融合孪生网络的目标跟踪算法  

Object tracking algorithm based on multilayer feature adaptive fusion siamese network

作　　者：董潇毅 王耀力 常青 孙永明[2] DONG Xiaoyi;WANG Yaoli;CHANG Qing;SUN Yongming(College of Information and Computer,Taiyuan University of Technology,Jinzhong 030600,China;Shanxi Academy of Forestry Sciences,Taiyuan 030000,China)

机构地区：[1]太原理工大学信息与计算机学院,山西晋中030600 [2]山西省林业科学研究院,山西太原030000

出　　处：《电子设计工程》2021年第22期102-107,113,共7页Electronic Design Engineering

基　　金：国家自然科学基金资助项目(61828601);山西省自然科学基金资助项目(201903D321003,201801D3121141)。

摘　　要：针对无人机在跟踪过程中存在目标尺度快速变化,跟踪易漂移、形变及相似干扰物等问题,提出了一种基于多层特征自适应融合孪生网络的目标跟踪算法。对孪生网络的CNN部分进行分析并改进,提出一种适用于无人机平台且表征能力更好的CNN模型;针对跟踪任务设计了一种由通道和空间注意力机制组成的双重注意力机制模块,将CNN输出的深浅层特征图组合输入到注意力机制模块后得到自适应融合特征图;将得到的自适应特征图通过区域生成网络(RPN)预测出目标在当前帧中的位置与尺度。通过在UAV123和DTB70无人机数据集进行测试,相较于基准孪生区域生成网络(SiamRPN),成功率分别提高了3.4%和5.8%,准确率提高了3.7%和6.1%;验证了该算法的有效性。In order to solve the problems such as easy loss of smallscale target tracking,deformation and similar semantic jammer in the tracking process of UAV,a target tracking algorithm based on multilayer feature adaptive fusion siamese network is proposed.The CNN part of the siamese network is analyzed and improved,and a CNN model which is suitable for UAV platform and has better characterization ability is proposed.A dual attention mechanism model composed of channel and spatial attention mechanism is designed for the tracking task.The deep and shallow feature maps of CNN output are combined into the attention mechanism module to obtain the adaptive fusion feature map.The fusion feature map is used to predict the location and scale of the target by using the region proposal network.By testing on the UAV123 and DTB70 UAV datasets,the success rate of the algorithm is improved by 3.4%and 5.8%,and the accuracy rate is improved by 3.7%and 6.1%compared with SiamRPN.Experiments show the effectiveness of the algorithm.

关 键 词：无人机 特征自适应融合 注意力机制 孪生网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]