摘要
DCFNet是一个轻量级的端到端的基于相关滤波的跟踪器,但未充分利用视频序列的时间和空间信息,难以应对场景复杂或目标变化较大的情景,针对DCFNet的该不足,提出一种新的采用时空采样网络采集特征结合相关滤波器的目标跟踪方法。该算法将可变形卷积层加入到时空采样网络,对附近历史帧的相关特征进行策略性采样,然后把采样得到的特征向量按照一定权重进行聚合,聚合后的特征向量送入相关滤波层,当跟踪到当前帧时通过定位滤波器的最大响应来估计目标的位置。分别在数据集OTB2013、OTB2015进行算法验证,结果表明,该算法较当前几类主流跟踪算法在跟踪成功率和精度上均有所提升。
DCFNet is an end-to-end lightweight network architecture,lacking video sequence and spatiotemporal information.In this work,we present a spatiotemporal sampling networks to achieve the real-time object tracking combined with the correlation filter.We use deformable convolution across space and time to leverage temporal information for visual tracking,which presenting an end-to-end lightweight network architecture.Adding the deformable convolution layer to the spatiotemporal sampling network,the networks make a strategic sampling of the relevant features from the historical frame,capturing a series of feature tensor.Finally,the feature tensors are aggregated according to weight w,aggregated feature tensors are fed into the CF layer.When tracking to the current frame,the position of the object is estimated by the maximum corresponding of the filter.Extensive experiments are performed on two challenging tracking datasets:OTB2013 and OTB2015,and the benchmark results show that the our networks perform well in speed and precision.
作者
谢颍晓
蔡敬菊
张建林
Xie Yingxiao;Cai Jingju;Zhang Jianlin(Institute of Optics and Electronics,Chinese Academy of Sciences,Chengdu 610209,China;University of Chinese Academy of Sciences,Beijing 100049,China)
出处
《国外电子测量技术》
2020年第1期88-92,共5页
Foreign Electronic Measurement Technology
关键词
时空采样网络
可变形卷积
相关滤波器
目标跟踪
深度学习
spatial-temporal network
deformable convolution
correlation filters
visual tracking
deep learning
作者简介
谢颍晓,硕士研究生,主要研究方向为深度学习、目标跟踪。E-mail:xieyingxiao17@mails.ucas.ac.cn;蔡敬菊,硕士生导师,副研究员,主要研究方向为成像探测和图像处理,包括可见光和红外的成像探测、多场景下各种类型目标的检测、识别和跟踪以及系统集成。E-mail:xueman1999@163.com;张建林,博士生导师,研究员,主要研究方向为智能图像处理计算机视觉研究,重点以机器学习为基础进行海量图像数据的分析理解,实现机器的智能化、自动化。E-mail:jlin_zh@163.com
