摘要
目标跟踪技术在智能监控、人机交互、无人驾驶等诸多领域得到了广泛的应用。近年来,学者们提出了许多高效的算法。然而,随着跟踪环境越来越复杂,目标跟踪算法在遮挡、光照变化、背景干扰等复杂环境下仍然面临着巨大的挑战,从而导致目标跟踪失败。针对上述问题,提出了一种基于时空注意力机制的目标跟踪算法。首先,采用孪生网络架构来提高对特征的判别能力;然后,引入改进的通道注意力机制和空间注意力机制,对不同通道和空间位置的特征施加不同的权重,并着重关注空间位置和通道位置上对目标跟踪有利的特征。此外,还提出了一种高效的目标模板在线更新机制,将第一帧图像特征与后续跟踪图像帧中置信度较高的图像特征进行融合,以降低发生目标漂移的风险。最后,在OTB2013和OTB2015数据集上对所提跟踪算法进行了测试。实验结果表明,所提算法的性能相比当前主流的跟踪算法提高了6.3%。
Object tracking technology is widely used in intelligent monitoring,human-computer interaction,unmanned driving and many other fields.In recent years,many efficient tracking methods are proposed.However,object tracking methods still face great challenges in the complex scenario such as occlusion,illumination variations,background clutter,which leads to tracking failure.To solve the above mentioned problems,in this paper,an effective object tracking algorithm is proposed based on temporal-spatial attention mechanism.Firstly,we utilize the Siamese network architecture to improve the discriminative ability of object features.Then,the improved channel attention module and spatial attention module are introduced into the Siamese network,which imposes different weights on the features of different channels and spatial positions and focuses on the features that are beneficial to object tracking in spatial and channel positions.In addition,an efficient online object template updating mechanism is developed,which combines the features of the first frame and the features of the following frames with high confidence to reduce the risk of the object drift.Finally,the proposed tracking algorithm is tested on OTB2013 and OTB2015 benchmarks.Experimental results show that the performance of the proposed algorithm improves by 6.3%compared with the current mainstream tracking algorithms.
作者
程旭
崔一平
宋晨
陈北京
郑钰辉
史金钢
CHENG Xu;CUI Yi-ping;SONG Chen;CHEN Bei-jing;ZHENG Yu-hui;SHI Jin-gang(School of Computer and Software,Nanjing University of Information Science and Technology,Nanjing 210044,China;Engineering Research Center of Digital Forensics,Ministry of Education,Nanjing University of Information Science and Technology,Nanjing 210044,China;School of Software Engineering,Xi’an Jiaotong University,Xi’an 710049,China)
出处
《计算机科学》
CSCD
北大核心
2021年第4期123-129,共7页
Computer Science
基金
国家自然科学基金(61802058,61911530397,62072251)
中国博士后科学基金项目(2019M651650)
南京信息工程大学人才启动经费(2018r057)。
关键词
深度学习
目标跟踪
孪生网络
注意力机制
模板更新
Deep learning
Object tracking
Siamese network
Attention mechanism
Template update
