Journal of Frontiers of Computer Science and Technology ›› 2020, Vol. 14 ›› Issue (5): 848-860.DOI: 10.3778/j.issn.1673-9418.1901063

Previous Articles     Next Articles

Single Target Tracking Algorithm Based on Multi-Fuzzy Kernel Fusion

CHEN Chen, DENG Zhaohong, GAO Yanli, WANG Shitong   

  1. 1. School of Digital Media, Jiangnan University, Wuxi, Jiangsu 214122, China
    2. Jiangnan Institute of Computing Technology, Wuxi, Jiangsu 214083, China
  • Online:2020-05-01 Published:2020-05-08

多模糊核融合的单目标跟踪算法

陈晨邓赵红高艳丽王士同   

  1. 1. 江南大学 数字媒体学院,江苏 无锡 214122
    2. 江南计算技术研究所,江苏 无锡 214083

Abstract:

For the problem how to accurately and quickly locate targets in the current target tracking field, the core content of most popular trackers is to combine a kernel method to train a discriminative classifier to distinguish the target from the surrounding environment. For example, the kernel correlation filter algorithm (KCF) combines the Fourier transform with the kernel discriminant classifier to improve the speed of tracking, and the improved fuzzy kernel correlation filter (FKCF) algorithm introduced by Takagi-Sugeno-Kang fuzzy logic system (TSK-FLS) is used to improve the accuracy of tracking. Some improved KCF-based algorithms have proposed solutions to partial tracking problems. However, existing algorithms still have some room for improvement in improving accuracy. Aiming at this deficiency, based on FKCF, a new multi-fuzzy kernels correlation filter (MFKCF) is derived using multi-kernel fusion. MFKCF inherits the characteristics of high speed of KCF and high accuracy of FKCF and blurs the polynomial kernel and the Gaussian kernel. And it combines the fuzzified kernel function as a new target kernel function. Due to the above two improvements, the proposed algorithm is better than KCF and FKCF in the accuracy of tracking. KCF, FKCF and MFKCF are carried out on 30 randomly selected videos on 4 databases such as OTB50. The experimental results show that MFKCF performs well on the whole, accuracies of the MFKCF on 10 types of common attributes are all improved.

Key words: kernel method, discriminant classifier, Fourier transform, TSK fuzzy logic system, multi-kernel fusion

摘要:

针对当前目标跟踪领域中如何准确迅速地对目标进行定位的问题,大部分流行跟踪器的核心内容是结合核方法去训练一个判别分类器来区分目标和周围环境。例如核相关滤波器算法(KCF)将傅里叶变换与核化判别分类器相结合来提升跟踪速度,以及引入TSK模糊逻辑系统(TSK-FLS)的模糊核相关滤波器(FKCF)算法来提高跟踪精度。一些基于KCF的改进算法对部分跟踪难题提出了解决方案,但这些算法在精度方面仍有一定的提升空间。针对此,在FKCF的基础上,从多核融合的角度推导出了一种新的多模糊核相关滤波器(MFKCF)。MFKCF继承了KCF高速的以及FKCF高精度的特性,将多项式核与高斯核进行模糊化,并且融合模糊化后的核函数作为新的目标核函数。由于上述两项改进,使所提算法在跟踪精度方面比KCF与FKCF更好。将KCF算法、FKCF算法与MFKCF算法在OTB50等4个数据集上的30个随机选取的视频进行了实验,实验结果表明MFKCF算法总体表现良好,10项常见属性上的精度均有提升。

关键词: 核方法, 判别分类器, 傅里叶变换, TSK模糊逻辑系统, 多核融合