文章摘要
杨海清,许倩倩,唐怡豪,孙道洋.结合卷积神经网络多特征融合的相关滤波跟踪[J].高技术通讯(中文),2020,30(10):1085~1092
结合卷积神经网络多特征融合的相关滤波跟踪
  
DOI:doi:10.3772/j.issn.1002-0470.2020.10.011
中文关键词: 目标跟踪; 卷积神经网络(CNN); 相关滤波; 特征提取; 可靠性加权
英文关键词: target tracking, convolutional neural network (CNN), correlation filtering,feature extraction, reliability weighting
基金项目:
作者单位
杨海清  
许倩倩  
唐怡豪  
孙道洋  
摘要点击次数: 287
全文下载次数: 219
中文摘要:
      针对相关滤波跟踪中的多特征融合问题,本文提出了基于多通道相关滤波框架结合卷积神经网络(CNN)多特征融合的跟踪算法。首先引入梯度直方图和颜色名特征,利用传统的特征提取方法将提取的特征进行简单的矢量相加。然后采用在ImageNet上训练的卷积神经网络进行特征提取,使用conv5-4卷积层的输出作为特征,再分别训练各自的相关滤波器,对特征响应进行可靠性加权求和获得目标位置。最后,通过最大响应值和平均峰值相关能量的变化来判断是否更新模型。在标准测试集(OTB-100)上进行实验测试,与5种基于相关滤波的主流算法进行性能对比。实验结果表明,本文算法在光照变化、尺度变化及遮挡等复杂情况下的鲁棒性和跟踪精度都优于其他算法。
英文摘要:
      To solve the multi-feature fusion problem in correlated filter tracking, this paper proposes a tracking algorithm combined with multi-channel correlation filter framework and multi-feature fusion convolution neural network (CNN). Firstly, this work introduces the gradient histogram and color name features, and uses the traditional feature extraction method to let the extracted features are simply vector-added. Secondly, the convolution neural network trained on ImageNet is used to extract the feature and the output of the conv5-4 convolutional layer is used as a feature. Afterwards, this work trains the correlation filters respectively. The target position is obtained by the reliable weighted sum of the feature response. Finally, by the change of both the maximum response value and the average peak-to correlation energy, it considers whether or not to update the prediction model.Experiments are carried out on the object tracking benchmark (OTB-100), and compared with five mainstream algorithms based on correlation filtering. The experimental results show that the robustness and tracking accuracy of the proposed algorithm are superior to the other algorithms’ in the complex conditions of illumination changes, scale variation and occlusion.
查看全文   查看/发表评论  下载PDF阅读器
关闭

分享按钮