考虑注意力的无锚框孪生网络目标跟踪算法
SIAMESE NETWORK OBJECT TRACKING BASED ON ANCHOR FREE CONSIDERING ATTENTION
孙仕棚 1兰时勇1
作者信息
- 1. 四川大学视觉合成图形图像技术国防重点学科实验室 四川 成都 610065
- 折叠
摘要
孪生(Siamese)网络是解决视觉目标跟踪任务的一种重要方法.无填充孪生网络(SiamDW)的跟踪器采用区域推荐网络(RPN)来进行目标的定位,需要预先设置锚框的高宽比等超参数,不仅调参繁杂,而且跟踪的准确率较低.为解决此问题,提出一种考虑通道注意力且无锚框的孪生网络目标跟踪方法.该方法以SiamDW为基线,引入无填充的DenseNet来提取目标的特征;在通道特征拼接的时候加入通道注意力模块,以提高目标特征的表征力;在无锚框设计的时候,采用一种矩形范围的方式对正负样本进行划分.实验结果表明,在VOT2016和VOT2018 数据集上,该算法跟踪的准确率分别比基线算法提高3 百分点和6 百分点.
Abstract
Siamese network is an important method for visual object tracking.SiamDW uses RPN to locate the object,which needs to preset more hyperparameters such as height-width ratio of anchors.It is tedious and a bit inaccurate.To address this issue,this paper proposes a Siamese visual object tracking algorithm based on anchor free considering channel attention.On the basis of SiamDW,this method introduced DenseNet with no padding to extract object's feature,and added channel attention module in channel concatenation to improve the feature representation.A rectangular range method was used to discriminate positive and negative samples.Results show that compared with baseline(SiamDW),the AUC of our method was increased by 3 and 6 percentage points on VOT2016 and VOT2018 datasets respectively.
关键词
孪生网络/目标跟踪/无锚框设计/无填充/通道注意力Key words
Siamese network/Object tracking/Anchor free/No padding/Channel attention引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据