基于近端策略优化模板更新的实时目标跟踪方法
Visual tracking based on template updating with proximal policy optimization
孙愉亚 1龚声蓉 2钟珊 3周立凡 3范利1
作者信息
- 1. 东北石油大学计算机与信息技术学院,黑龙江大庆 163000
- 2. 东北石油大学计算机与信息技术学院,黑龙江大庆 163000;常熟理工学院计算机科学与工程学院,江苏苏州 215000
- 3. 常熟理工学院计算机科学与工程学院,江苏苏州 215000
- 折叠
摘要
基于孪生网络的 目标跟踪算法往往采用第一帧的外观特征作为固定模板,难以应对目标外观剧烈变化等问题.为此,所提算法在孪生网络的基础上,引入深度强化学习,将模板更新问题建模为马尔可夫决策过程,采用近端策略优化算法进行优化,减少因 目标外观变化带来的误差积累.针对孪生网络跟踪算法搜索域太小,无法全局搜索目标的问题,引入全局检测算法,找回丢失的 目标.所提跟踪算法能够自适应更新模板和全局检测丢失的 目标,在OTB数据集和GOT-10k数据集上进行测试,实验结果表明,该方法较代表性方法,具有实时性强和准确率高的优点,能够很好应对目标外观形变以及目标丢失.
Abstract
The siamese network based object tracking algorithm often uses the appearance features of the first frame as the fixed template,which is difficult to deal with the sharp changes in the appearance of the target.Therefore,based on the siamese net-work,the deep reinforcement learning was introduced,the template updating problem was modeled as a Markov decision process,and the proximal policy optimization was used for optimization,so as to reduce the error accumulation caused by the change of object appearance.To solve the problem that the search area of the siamese network tracking algorithm is too small to search the target globally,a global detection algorithm was introduced to recover the lost target.The template was updated adaptively and the missing target was detected globally.It was tested on OTB dataset and GOT-10k dataset.Experimental results show that the proposed method is more representative,has the advantages of strong real-time performance and high accu-racy,and can deal with the external deformation of the object and object loss well.
关键词
目标跟踪/深度强化学习/近端策略优化/马尔可夫决策过程/全局检测/更新模板/孪生网络Key words
object track/deep reinforcement learning/proximal policy optimization/Markov decision process/global detect/up-date template/siamese network引用本文复制引用
基金项目
国家自然科学基金面上项目(61972059)
国家自然科学基金面上项目(42071438)
国家自然科学基金面上项目(62102347)
中国博士后科学基金面上项目(2021M69236)
江苏省自然科学基金面上项目(BK20191474)
江苏省自然科学基金面上项目(BK20191475)
江苏省自然科学基金面上项目(BK20221403)
吉林大学符号计算与知识工程教育部重点实验室项目(93K172021K01)
吉林大学符号计算与知识工程教育部重点实验室项目(93K172017K18)
出版年
2024
NETL
NSTL
维普
万方数据