融合多模板注意力深度网的自适应目标框跟踪算法
Adaptive target box tracking algorithm by integrating multi-template attention deep network
仲训杲 1范东嘉 2仲训昱 3周承仙 2赵晶 1刘强4
作者信息
- 1. 厦门理工学院电气工程与自动化学院,福建厦门 361024;厦门市高端电力装备及智能控制重点实验室,福建厦门 361024
- 2. 厦门理工学院电气工程与自动化学院,福建厦门 361024
- 3. 厦门大学航空航天学院,福建厦门 361002
- 4. 牛津大学精神学系,牛津OX37JX
- 折叠
摘要
现有深度网络跟踪算法应对相似物体干扰、尺度变化、形变模糊、遮挡等问题存在挑战,为此提出一种融合多模板注意力机制的鲁棒深度网络算法.在SiamFc深度网络分支中构建通道和空间多模板注意力机制,以加强网络对目标特征的提取能力;融合浅层和深层卷积特征实现跟踪目标的精确聚焦,以克服相似物干扰问题;采用自适应回归网络学习目标采样点与目标边界之间的距离,实现目标区域的动态预测,有效应对目标尺度变化问题.另外,通过计算分类特征的APCE均值和最大值建立模板在线更新策略,实现网络自适应目标形变模糊与遮挡等问题.对OTB 100和VOT 2016等公开数据集的测试结果表明,与目前先进的SiamFc及改进方法相比,所提出算法在动态目标跟踪的准确率和成功率上均得到有效提升,具有强鲁棒性能.
Abstract
In view of the similar object interference,target scale changes,deformation blur,occlusion and other challenging problems for existing deep network tracking algorithms.This paper proposes a robust deep network tracking algorithm by integrating multi-template attention mechanism.The proposed method builds a channel and spatial multi-template attention mechanism in the branch of the Siamfc network,so as to strengthen the ability of the deep network for features extraction,and by integrates shallow and deep convolution features to achieve the accurate focus of tracking targets,so as to overcome the interference problem of similar objects.The adaptive regression network is used to learn the distance between the target sampling point and the target boundary,so as to realize the dynamic prediction of the target area and effectively deal with the problem of target scale change.In addition,the target template online update strategy is established by calculating the APCE mean value and maximum value of classification features,so as to realize the network adaptive the target deformation blur and occlusion problems.Through the test of OTB100,VOT2016 and other public data sets,the results show that compared with the current advanced deep network frameworks such as Siamfc and its improved method,the proposed algorithm has effectively improved the accuracy and success rate of dynamic target tracking,and the research method has a strong robust performance.
关键词
深度网络/目标跟踪/自适应框/注意力机制/模板更新Key words
deep network/object tracking/adaptive box/attention mechanism/template update引用本文复制引用
基金项目
国家自然科学基金项目(61703356)
福建省自然科学基金项目(2022J011256)
福建省自然科学基金项目(2020J01285)
厦门市青年创新基金项目(3502Z20206071)
出版年
2024
NETL
NSTL
万方数据