计算机工程与应用 ›› 2019, Vol. 55 ›› Issue (15): 161-168.DOI: 10.3778/j.issn.1002-8331.1812-0129

• 模式识别与人工智能 • 上一篇    下一篇

多通道特征和择优并行更新的核相关滤波跟踪

胡昭华,李高飞,陈胡欣   

  1. 1.南京信息工程大学 电子与信息工程学院,南京 210044
    2.南京信息工程大学 江苏省大气环境与装备技术协同创新中心,南京 210044
  • 出版日期:2019-08-01 发布日期:2019-07-26

Multi-Channel Feature and Preferred Parallel Update for Kernel Correlation Filter Tracking

HU Zhaohua, LI Gaofei, CHEN Huxin   

  1. 1.College of Electronics and Information Engineering, Nanjing University of Information Science and Technology, Nanjing 210044, China
    2.Jiangsu Collaborative Innovation Center on Atmospheric Environment and Equipment Technology, Nanjing University of Information Science and Technology, Nanjing 210044, China
  • Online:2019-08-01 Published:2019-07-26

摘要: 针对传统核跟踪算法单一特征的局限性、目标模板和特征外观模板更新的不足,提出了一种多通道特征和择优并行更新的核相关滤波跟踪算法。采用多通道特征提取方式:上支路采用卷积神经网络提取深度特征,下支路则将HOG特征和CN特征相结合用于训练与跟踪。采用新的目标模板和特征外观模板更新方式:择优并行更新,取不同支路当前帧的最大响应值作为最佳目标位置,下一帧中两个支路的模板更新采用前一帧最优位置的参数同时进行更新,直到跟踪结束,多支路的择优并行更新弥补了单一支路更新的不足。实验表明该算法能在不同挑战因子下实现更加鲁棒的跟踪过程。

关键词: 核相关滤波, 目标模板, 多通道特征, 择优并行更新, 卷积神经网络

Abstract: Aiming at the limitations of single features and insufficiencies update of target templates and feature appearance templates intraditional kernel tracker, a multi-channel feature and preferred parallel update for kernel correlation filter tracking is proposed. The multi-channel feature extraction method is adopted which means the upper branch uses the convolutional neural network to extract the depth features, and the lower branch combines the features of HOG and CN for training and tracking. A novel preferred parallel update method is proposed for updating the target templates and feature appearance templates. In the current frame, the maximum response value in both branches is considered as the optimal target position. In the next frame, the templates of the two branches are updated simultaneously with the parameters of the optimal position of the previous frame until the end of the tracking. The optimal parallel update of multiple branches makes up for the deficiency of single branch update. Experiments show that this algorithm can achieve more robust tracking result under different challenge factors.

Key words: kernel correlation filtering, target template, multi-channel feature, preferred parallel update, convolutional neural network