删除或更新信息,请邮件至freekaoyan#163.com(#换成@)

孪生网络框架下融合显著性和干扰在线学习的航拍目标跟踪算法

本站小编 Free考研考试/2022-01-03

孙锐,
方林凤,,
梁启丽,
张旭东
1.合肥工业大学计算机与信息学院 合肥 230009
2.工业安全与应急技术安徽省重点实验室 合肥 230009
基金项目:国家自然科学基金(61471154, 61876057),安徽省重点研发计划-科技强警专项(202004d07020012)

详细信息
作者简介:孙锐:男,1976年生,教授,主要研究方向为计算机视觉与机器学习
方林凤:女,1994年生,硕士生,研究方向为图像信息处理和计算机视觉
梁启丽:女,1995年生,硕士生,研究方向为图像信息处理和计算机视觉
张旭东:男,1966年生,教授,主要研究方向为智能信息处理
通讯作者:方林凤 f_linf@163.com
中图分类号:TN911.73; TP391

计量

文章访问数:611
HTML全文浏览量:132
PDF下载量:91
被引次数:0
出版历程

收稿日期:2020-03-03
修回日期:2020-10-21
网络出版日期:2020-11-19
刊出日期:2021-05-18

Siamese Network Combined Learning Saliency and Online Leaning Interference for Aerial Object Tracking Algorithm

Rui SUN,
Linfeng FANG,,
Qili LIANG,
Xudong ZHANG
1. School of Computer and Information, Hefei University of Technology, Hefei 230009, China
2. Anhui Province Key Laboratory of Industry Safety and Emergency Technology, Hefei 230009, China
Funds:The National Natural Science Foundation of China (61471154, 61876057), The Key Research Plan of Anhui Province - Strengthening Police with Science and Technology (202004d07020012)


摘要
摘要:针对一般跟踪算法不能很好地解决航拍视频下目标分辨率低、视场大、视角变化多等特殊难点,该文提出一种融合目标显著性和在线学习干扰因子的无人机(UAV)跟踪算法。通用模型预训练的深层特征无法有效地识别航拍目标,该文跟踪算法能根据反向传播梯度识别每个卷积滤波器的重要性来更好地选择目标显著性特征,以此凸显航拍目标特性。另外充分利用连续视频丰富的上下文信息,通过引导目标外观模型与当前帧尽可能相似地来在线学习动态目标的干扰因子,从而实现可靠的自适应匹配跟踪。实验证明:该算法在跟踪难点更多的UAV123数据集上跟踪成功率和准确率分别比孪生网络基准算法高5.3%和3.6%,同时速度达到平均28.7帧/s,基本满足航拍目标跟踪准确性和实时性需求。
关键词:目标跟踪/
无人机航拍场景/
孪生网络/
目标显著性/
在线学习干扰因子
Abstract:In view of the fact that the general tracking algorithm can not solve the special problems such as low resolution, large field of view and many changes of view angle, a Unmanned Aerial Vehicle (UAV) tracking algorithm combining target saliency and online learning interference factor is proposed. The deep feature that the general model pre-trained can not effectively identify the aerial target, the tracking algorithm can better select the salient feature of each convolution filter according to the importance of the back propagation gradient, so as to highlight the aerial target feature. In addition, it makes full use of the rich context information of the continuous video, and learn the interference factor of the dynamic target online by guiding the target appearance model as similar as possible to the current frame, so as to achieve reliable adaptive matching tracking. It is proved that the tracking success rate and accuracy rate of the algorithm are 5.3% and 3.6% higher than that of the siamese network benchmark algorithm on the more difficult UAV123 dataset, respectively, and the speed reaches an average of 28.7 frames per second, which basically meet the aerial target tracking accuracy and real-time requirements.
Key words:Object tracking/
Unmanned Aerial Vehicle (UAV) scene/
Siamese network/
Target saliency/
Online leaning interference factor



PDF全文下载地址:

https://jeit.ac.cn/article/exportPdf?id=f28b8e95-8562-4046-9fb5-560e237ed90f
相关话题/计算机 干扰 网络 信息 视觉