摘要
本文是在深度特征与相关滤波相结合的高效卷积运算符(ECO)目标跟踪算法基础上进行的改进。首先,为了提高跟踪速度,提出“浅层特征不插值,深层特征插值”的卷积神经网络(CNN)分层插值处理方法,对具有较高分辨率的浅层特征不插值,对分辨率低的深层特征进行插值计算来提高分辨率;其次,改进了样本空间分类策略,给CNN特征层分配不同的权重,突出不同特征层对样本间距离的影响,并且将所有样本信息都保留在训练样本集中;最后,应用判别尺度空间跟踪(DSST)算法提出的对目标尺度估计的方法,增加了目标尺度的候选数量,使尺度估计更加准确。实验结果验证了所设计算法的有效性。
This paper presents improvement of efficient convdution operator(ECO)which combinates depth features and correlation filters.Firstly,in order to improve the speed of tracker,‘shallow feature no interpolation,deep feature interpolation’is proposed for convolutional neural networks(CNN)hierarchical interpolation,i.e.the shallow features with higher resolution are not interpolated,but the deep features with low resolution are interpolated to promote resolution.Secondly,the sample space classification strategy is improved,different weights is assigned for CNN feature layer,the influence of different feature layers on the distance between samples is highlighted,and all sample information is retained in the training sample set.Finally,the discriminative scale space tracker(DSST)is used to estimate the target scale,and the number of reference target scale is increased,so the target scale is estimated accurately.The experimental results verified the effectiveness of the designed algorithm.
作者
陈志旺
王昌蒙
王莹
宋娟
彭勇
Chen Zhiwang;Wang Changmeng;Wang Ying;Song Juan;Peng Yong(Key Laboratory of Industrial Computer Control Engineering of Hebei Province,Yanshan University,Qinhuangdao 066004;National Engineering Research Center for Equipment and Technology of Cold Strip Rolling,Yanshan University,Qinhuangdao 066004)
出处
《高技术通讯》
EI
CAS
北大核心
2020年第6期570-578,共9页
Chinese High Technology Letters
基金
国家自然科学基金(61573305)资助项目。