视频单目标跟踪研究进展综述  

Single Object Tracking Research:A Survey

作　　者：韩瑞泽 冯伟[1,2] 郭青 胡清华[1] HAN Rui-Ze;FENG Wei;GUO Qing;HU Qing-Hua(College of Intelligence and Computing,Tianjin University,Tianjin 300350;Key Research Center for Surface Monitoring and Analysis of Cultural Relics,Tianjin 300350)

机构地区：[1]天津大学智能与计算学部,天津300350 [2]国家文物局文物本体表面监测与分析研究重点科研基地,天津300350

出　　处：《计算机学报》2022年第9期1877-1907,共31页Chinese Journal of Computers

基　　金：天津市自然科学基金项目(18JCYBJC15200);天津市研究生科研创新项目(2021YJSB174);国家自然科学基金项目(U1803264,62072334)资助

摘　　要：视频目标跟踪是计算机视觉中的重要任务之一,在实际生活中有着广泛的应用,例如视频监控、视觉导航等.视频目标跟踪任务也面临着诸多挑战,如目标遮挡、目标形变等情形.为解决目标跟踪中的挑战,实现精确高效的目标跟踪,近年来出现大量的目标跟踪算法.本文介绍了近十年来视频目标跟踪领域两大主流算法框架(基于相关滤波和孪生网络的目标跟踪算法)的基本原理、改进策略和代表性工作,之后按照网络结构分类介绍了其他基于深度学习的目标跟踪算法,还从解决目标跟踪所面临挑战的角度介绍了应对各类问题的典型解决方案,并总结了视频目标跟踪的历史发展脉络和未来发展趋势.本文还详细介绍和比较了面向目标跟踪任务的数据集和挑战赛,并从数据集的数据统计和算法的评估结果出发,总结了各类视频目标跟踪算法的特点和优势.针对目标跟踪未来发展趋势,本文认为视频目标跟踪还面临诸多难题亟需解决,例如当前的算法往往无法在长时间、低功耗、抗干扰的环境下实地应用.未来,考虑多模态数据融合,如将深度图像、红外图像数据与传统彩色视频联合分析,将会为目标跟踪带来更多新的解决方案.目标跟踪任务也将会和其他任务,如视频目标检测、视频目标分割,相互促进共同发展.Visual object tracking is an important and fundamental task in computer vision,which has many real-world applications,e.g.,video surveillance,visual navigation and robotic service.Visual object tracking also has many challenges,such as object loss,object deformation,background clutters,and object fast motion.To solve the above problems and track the target accurately and efficiently,many visual object tracking algorithms have been emerged in recent years.In this paper,we first review the two most popular tracking frameworks in the past ten years,i.e.,the Correlation Filter(CF)and Siamese network based visual object tracking.We present the rationale,the improvement strategy,and the representative works of the above two frameworks in detail.Specifically,the CF technology has been used in visual object tracking for over ten years,which has a good balance between the tracking accuracy and running speed.In CF tracking,the target is located by applying a circular convolution operation on the learned filter and the current frame,which can be efficiently achieved by the Fast Fourier Transform(FFT).The Siamese network based trackers locate the target from the candidate patches through a matching function offline learned on abundant training data in terms of image pairs.The matching function is modeled by a two-branch convolutional neural network(CNN)with shared parameters to learn the similarity between the target and the candidate patches.Besides the above two frameworks,we then present some other deep learning based tracking methods categorized by different network structures,e.g.,RNN(Recurrent Neural Network),GCN(Graph Convolutional Network),etc.We also introduce some classical strategies for handling the challenges in the visual object tracking problem.From the recent tracking methods,we find that the development direction of the methods shows a diversified trend.More new network structures and skills have been applied to object tracking task.Although other deep tracking methods show a diversified trend,they have not

关 键 词：视频目标跟踪 相关滤波跟踪算法 孪生网络跟踪算法 目标跟踪数据集 目标跟踪发展 综述 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]