深度视觉目标跟踪进展综述  

Recent advance in deep visual object tracking

作　　者：王宁[1] 席茂 周文罡[1] 李礼[1] 李厚强[1] Wang Ning;Xi Mao;Zhou Wengang;Li Li;Li Houqiang(MOE-Microsoft Key Laboratory of Multimedia Computing and Communication,University of Science and Technology of China,Hefei 230027,China)

机构地区：[1]中国科学技术大学多媒体计算与通信教育部-微软联合实验室,安徽合肥230027

出　　处：《中国科学技术大学学报》2021年第4期335-344,共10页JUSTC

基　　金：中国科学技术大学青年创新重点基金(YD3490002001)资助.

摘　　要：视频目标跟踪是计算机视觉领域的一个重要研究课题.近年来,随着深度学习在视觉目标跟踪领域获得了巨大的成功,一系列优秀的深度跟踪算法涌现出来.本文回顾了近年来深度目标跟踪领域的进展.首先,我们详细讨论了近十年来跟踪领域数据集的发展趋势,这些数据集不仅全面地评估了算法性能同时为模型训练提供了极大的便利.其次,我们分类讨论了几大类经典的深度学习跟踪框架,包括深度相关滤波器跟踪、分类式网络跟踪、双路网络跟踪、基于梯度的深度跟踪算法以及基于Transformer的跟踪算法.最后,我们对全文内容进行总结,并指出未来的发展趋势.Visual object tracking is an important branch in computer visions.In recent years,with the remarkable success of deep learning techniques,a series of deep tracking algorithms have emerged with impressive performances.In this paper,we review the recent development of deep learning based trackers.First,we revisit the development of tracking benchmarks in the last decade.These tracking datasets not only comprehensively help evaluate the tracking algorithms but also largely support the model training of deep trackers.Next,we discuss several representative tracking frameworks including deep correlation filter tracking,classification-based tracking networks,Siamese tracking networks,gradient-based tracking networks and Transformer based deep trackers.Finally,we conclude the paper and discuss the potential future research directions of the visual tracking.

关 键 词：深度目标跟踪 跟踪数据集 相关滤波器 分类式跟踪网络 双路跟踪网络 梯度跟踪网络 

分 类 号：P595[天文地球—地球化学]