基于改进孪生网络的无人机遥感目标跟踪  

Target tracking in UAV remote sensing videos based on improved Siamese network

作　　者：窦凯奇 朱福珍[1] DOU Kaiqi;ZHU Fuzhen(College of Electronic Engineering,Heilongjiang University,Harbin 150080,China)

机构地区：[1]黑龙江大学电子工程学院,哈尔滨150080

出　　处：《黑龙江大学自然科学学报》2022年第6期750-756,共7页Journal of Natural Science of Heilongjiang University

基　　金：国家自然科学基金资助项目(61601174);黑龙江省博士后科研启动基金资助项目(LBH-Q17150);黑龙江省普通高等学校电子工程重点实验室(黑龙江大学)开放课题及省高校科技创新团队资助项目(2012TD007);黑龙江省省属高等学校基本科研业务费资助项目(KJCXZD201703);黑龙江省自然科学基金资助项目(F2018026)。

摘　　要：为了实现无人机(Unmanned aerial vehicle,UAV)视频中遥感目标的精确跟踪,对传统基于孪生网络的跟踪架构进行了改进。使用改进的ResNet-50作为骨干网络,实现模板和搜索区域进行深度特征提取,以提升骨干网络对于深度特征的提取能力,将孪生骨干网络所提取的特征通过深度互相关操作进行融合。为了抑制低质量预测边界框的生成,本研究在预测头部网络中聚合了一个质量评估分支。对于深度互相关操作得到的特征响应图,直接以无锚框的方式在统一的全卷积网络中预测目标的位置以及纵横比。UAV123数据集上的试验结果表明,相较于传统算法,所设计的跟踪算法在精确度曲线和成功率曲线的得分分别提升了2.2%和4.7%,能够很好地处理快速运动、光照变化以及尺度变化等情况。In order to achieve accurate tracking in Unmanned aerial vehicle(UAV)remote sensing videos,the traditional tracking framework is improved based on Siamese network.The modified ResNet-50 is used as the backbone network to extract depth features from the template and search region,increasing the feature extraction capability of the network.These extracted features are fused by depth cross-correlation operation.In order to remove low-quality prediction bounding boxes,a quality assessment branch is integrated into the prediction head.For these feature response maps obtained by depth feature cross-correlation operation,the position and scale of the target object are predicted directly in a unified fully convolutional network by an anchor-free manner.These experimental results on UAV123 dataset show that,compared to traditional algorithm,our proposed tracking algorithm scores 2.2%and 4.7%better on the precision and success rate,respectively,and is able to handle fast motion,illumination variation and scale variation well.

关 键 词：无人机 孪生网络 目标跟踪 遥感视频 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]