计算机工程 ›› 2020, Vol. 46 ›› Issue (2): 279-285.doi: 10.19678/j.issn.1000-3428.0053954

• 图形图像处理 • 上一篇    下一篇

基于增强语义与多注意力机制学习的深度相关跟踪

周双双1,2, 宋慧慧1,2, 张开华1,2, 樊佳庆1,2   

  1. 1. 南京信息工程大学 自动化学院, 南京 210044;
    2. 江苏省大数据分析技术重点实验室, 南京 210044
  • 收稿日期:2019-02-19 修回日期:2019-04-04 发布日期:2020-02-12
  • 作者简介:周双双(1988-),男,硕士研究生,主研方向为目标跟踪;宋慧慧、张开华,教授、博士;樊佳庆,硕士研究生。
  • 基金项目:
    国家自然科学基金(61872189,61876088);江苏省自然科学基金(BK20170040)。

Deep Correlation Tracking Based on Reinforced Semantic and Multi-attention Mechanism Learning

ZHOU Shuangshuang1,2, SONG Huihui1,2, ZHANG Kaihua1,2, FAN Jiaqing1,2   

  1. 1. School of Automation, Nanjing University of Information Science and Technology, Nanjing 210044, China;
    2. Jiangsu Key Laboratory of Big Data Analysis Technology, Nanjing 210044, China
  • Received:2019-02-19 Revised:2019-04-04 Published:2020-02-12

摘要: 在基准可判别相关滤波器网络(DCFNet)目标跟踪过程中,遮挡和运动模糊情况会导致目标发生漂移。针对该问题,结合增强语义与多注意力机制深度学习,设计一种端到端的相关滤波器网络RACFNet。由EDNet网络得到高级语义信息弥补单独低级特征表示的不足,同时加入通道和空间残差注意力机制,使网络能够对不同的跟踪对象提取出更具针对性的表观信息。在此基础上,通过添加相关滤波层并输出响应图最大值推测目标位置。在OTB-2013和OTB-2015基准测试集上的实验结果表明,RACFNet跟踪速度平均可达92帧/s,跟踪成功率较DCFNet分别提高8.20%和10.69%。

关键词: 增强语义, 注意力机制, 相关滤波, 傅里叶域计算, 目标跟踪

Abstract: In baseline tracker Discriminant Correlation Filter Network(DCFNet),occlusion and motion blur often cause target drift.To address the problem,this paper proposes a correlation filtering network of end to end structure,RACFNet.First,EDNet is used to obtain high-level semantic information to make up for low-level feature representation.Then both channel-wise and spatial residual attention mechanism are introduced to enable the network to extract more specific information for different tracking objects.At last,correlation filters are utilized to estimate the target locations according to the output maximum response value.Experimental results on OTB-2013 and OTB-2015 benchmarking datasets show that RACFNet runs at an average speed of 92 frame/s.The success rate of tracking is improved by 8.20% and 10.69% compared with DCFNet.

Key words: reinforced semantic, attention mechanism, correlation filtering, Fourier domain calculation, object tracking

中图分类号: