计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (7): 210-220.DOI: 10.3778/j.issn.1002-8331.1910-0317

• 图形图像处理 • 上一篇    下一篇

快速尺度估计的增强型多核相关滤波算法

杨佳霖,王文伟,熊晓璇,和世瑛   

  1. 武汉大学 电子信息学院,武汉 430072
  • 出版日期:2020-04-01 发布日期:2020-03-28

Fast Scale Estimation Based on Enhanced Multi-kernel Correlation Filter Algorithm

YANG Jialin, WANG Wenwei, XIONG Xiaoxuan, HE Shiying   

  1. School of Electronic Information, Wuhan University, Wuhan 430072, China
  • Online:2020-04-01 Published:2020-03-28

摘要:

针对核相关滤波(KCF)跟踪算法在复杂环境下其定位性能和稳定性差的问题,提出了一种快速尺度估计的增强型多核相关滤波跟踪算法。该算法针对核相关滤波算法无法适应跟踪过程中目标尺度变化,将快速判别式尺度估计移植至核相关滤波跟踪框架,解决了跟踪过程的目标尺度问题。对于单个特征的单核相关滤波器在复杂环境中跟踪适应性差的问题,提出了一种多特征互补的多核相关滤波器。该滤波器利用KCF多通道特性以及不同特征可以描述不同信息,采用多个相同内核的线性组合,每个内核对应一个特征,并结合快速尺度估计,在保证算法实时性的同时进一步提高跟踪性能。通过在OTB2013目标跟踪数据集上进行实验,该算法与近年来性能优异的算法进行对比,结果表明,与传统的使用HOG特征的KCF算法相比精度上提高了10.9%,成功率提高了16.2%;与使用CN特征的CN2算法相比,精度上提高了20.6%,成功率提高了19.6%。实验结果表明,所提算法在目标尺度变化以及复杂环境下的跟踪效果均优于其余相关滤波算法,证明了该算法的有效性以及鲁棒性。

关键词: 目标跟踪, 核相关滤波, 多特征互补, 多核学习

Abstract:

Aiming at the problem that the Kernel Correlation Filtering(KCF) tracking algorithm has poor positioning performance and stability in complex environments, an enhanced multi-core correlation filtering tracking algorithm for fast scale estimation is proposed. The algorithm is aimed at the problem that the kernel correlation filtering algorithm can not adapt to the target scale change in the tracking process. The fast discriminant scale estimation is transplanted to the kernel correlation filter tracking framework, and the target scale problem of the tracking process is solved. A multi-core correlation filter with multiple features is proposed for the problem that the single-core correlation filter of single feature has poor tracking adaptability in complex environment. The filter can describe different information by using KCF multi-channel characteristics and different features. It adopts linear combination of multiple identical cores, each core corresponds to one feature, and combines fast scale estimation to improve the tracking performance while ensuring the real-time performance of the algorithm. By conducting experiments on the OTB2013 target tracking dataset and comparing with the algorithms with excellent performance in recent years, the results show that compared with the traditional KCF algorithm using HOG features, the accuracy is improved by 10.9% and the success rate is increased by 16.2%. Compared with the CN2 algorithm using the CN feature, the accuracy is improved by 20.6% and the success rate is increased by 19.6%. The experimental results show that the proposed algorithm is better than the other related filtering algorithms in the target scale change and complex environment, which proves the effectiveness and robustness of the proposed algorithm.

Key words: object tracking, kernelized correlation filter, multiple feature complementation, multi-kernel learning