判别式相关滤波器的目标跟踪综述  

Survey of Visual Tracking Based on Discriminative Correlation Filters

作　　者：徐宁 王娟娟 郭晓雨 赵增顺 XU Ning;WANG Juan-juan;GUO Xiao-yu;ZHAO Zeng-shun(Institute of Electronic Information Engineering,Shandong University of Science and Technology,Qingdao 266590,China)

机构地区：[1]山东科技大学电子信息工程学院,山东青岛266590

出　　处：《小型微型计算机系统》2020年第12期2484-2493,共10页Journal of Chinese Computer Systems

摘　　要：目标跟踪是计算机视觉领域的重要研究方向,单目标跟踪主要分为深度学习、判别式相关滤波器和传统方法.得益于频域计算的高效性,本文选择以判别式相关滤波器为切入点.首先介绍了判别式相关滤波器进行目标跟踪的原理,然后围绕着基础框架进行横向展开,解决边界效应成为判别式相关滤波器发展的分水岭.伴随着卷积神经网络的发展,描述了基于预训练模型的特征提取和深度学习与相关滤波器框架相结合的两个方向,最后总结了关于判别式相关滤波器的发展脉络图.对比于特征插值与置信图融合的方向,在模型中构建合理的约束项进行模型创新成为一个重要方向.实验部分呈现了跟踪器在OTB-2015与VOT-2018数据集下的对比结果与排名,并进行简短的分析.Visual tracking is an important research direction of computer vision,single target tracking mainly divided into deep learning,discriminative correlation filters and traditional methods.Benefiting from the efficiency of frequency domain calculations,this paper chooses to select discriminative correlation filters(DCF)as the main point.Firstly,this paper introduces the principle of DCF for visual tracking,and then expands horizontally around the basic framework,to solve the boundary effect becomes a watershed of DCF.With the development of convolutional neural networks,there exist tw o directions of DCF,feature extraction based on pre-trained model and deep learning combined with the DCF framework.Finally,we summarize an evolution chart of DCF.Compared with the direction of feature interpolation and confidence map fusion,to construct reasonable constraints in the model for model innovation becomes an important direction.The experiment presents the comparison results and ranking of trackers under the OTB-2015 and VOT-2018 datasets,and makes a short analysis.

关 键 词：目标跟踪 相关滤波器 判别式相关滤波器 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]