残差深度特征和漂移检测的核相关滤波跟踪
Residual depth feature and drift detection for kernel correlation filter tracking
摘要点击 71  全文点击 90  投稿时间:2018-04-25  修订日期:2018-09-26
查看全文  查看/发表评论  下载PDF阅读器
DOI编号  10.7641/CTA.2018.80299
  2019,36(4):613-621
中文关键词  目标跟踪  深度神经网络  残差融合特征  检测模型漂移策略  模型更新
英文关键词  target tracking  deep neural network  residual fusion feature  detection model drift strategy  model update
基金项目  国家自然科学基金项目(61601230),江苏省自然科学基金项目(BK20141004)
学科分类代码  
作者单位E-mail
胡昭华 南京信息工程大学 zhaohua_hu@163.com 
郑伟 南京信息工程大学  
钱坤 南京信息工程大学  
中文摘要
      传统特征的片面性,传统跟踪模型对于模型漂移问题检测手段和补救措施的缺乏,限制着传统跟踪算法的性能。因此,本文提出一种漂移检测的残差深度特征目标跟踪算法,通过卷积神经网络提取分层特征,然后在卷积神经网络加入残差结构,连接不同的网络层,实现浅层和深层特征的融合,不需要人为设计特征融合方式,网络结构能够自动实现特征融合的功能,用深度特征区分目标和背景,比传统特征更具有分辨力;在预测当前帧的目标位置时,提出了一个检测模型漂移的策略,设计了一个响应强度下降计数器,通过对比相邻帧响应强度的大小计数,根据计数器的数值,用来判断是否出现模型漂移,以采取相对应的模型更新方案作为补救措施,实现精确跟踪。在与当下的几种跟踪算法比较中,本文提出的跟踪算法在跟踪精度和鲁棒性都优于所对比的算法。
英文摘要
      The one-sidedness of the traditional features, the lack of detection methods and remedial measures of the traditional tracking model for model drift problems all limit the performance of traditional tracking algorithms. Therefore, this paper proposes a drift detection residual fusion feature target tracking algorithm, which extracts hierarchical features through convolutional neural network, then adds residual structure in the convolutional neural network, connects different network layers, and realizes shallow and deep features. The fusion does not require artificially designed feature fusion methods. The network structure can automatically realize the features of feature fusion, distinguishing targets and backgrounds with depth features, and having more resolution than traditional features; when predicting the target position of the current frame, a detection is proposed. The strategy of drifting the model is to design a response intensity reduction counter, which is used to compare the magnitude of the response strength of adjacent frames. According to the value of the counter, it is used to judge whether there is a model drift, and a corresponding model update scheme is taken as a remedial measure to realize accurate tracking. In comparison with several current tracking algorithms, the tracking algorithm proposed in this paper is superior to the compared algorithm in tracking accuracy and robustness.