Abstract—Thecore component of most modern trackers is a discriminative classifier, taskedwith distinguishing between the target and the surrounding environment. To copewith natural image changes, this classifier is typically trained withtranslated and scaled sample patches. Such sets of samples are riddled withredundancies – any overlapping pixels are constrained to be the same. Based onthis simple observation, we propose an analytic model for datasets of thousandsof translated patches. By showing that the resulting data matrix is circulant,we can diagonalize it with the Discrete Fourier Transform, reducing bothstorage and computation by several orders of magnitude. Interestingly, forlinear regression our formulation is equivalent to a correlation filter, usedby some of the fastest competitive trackers. For kernel regression, however, wederive a new Kernelized Correlation Filter (KCF), that unlike other kernelalgorithms has the exact same complexity as its linear counterpart. Building onit, we also propose a fast multi-channel extension of linear correlationfilters, via a linear kernel, which we call Dual Correlation Filter (DCF). BothKCF and DCF outperform top-ranking trackers such as Struck or TLD on a 50videos benchmark, despite running at hundreds of frames-per-second, and beingimplemented in a few lines of code (Algorithm 1). To encourage furtherdevelopments, our tracking framework was made open-source.
IndexTerms—Visual tracking, circulant matrices, discrete Fourier transform, kernel methods,ridge regression, correlation filters.
摘要
很多当代跟踪器的核心部分是识别分类器，他的任务是区分目标和目标周周环境；对于图片的自然改变，分类器使用变换缩放的结果用来训练；这样的样本集合充斥着冗余，任何重合像素被被限制为相同；基于这个简单的观察，我们提出了一个对于成千上万的训练块的分析模型；通过证明数据矩阵是一个循环矩阵，我们可以通过离散傅里叶变换对角化循环矩阵，以此减少几个数量级的存储和计算；有趣的是，对于线性滤波，我们的公式相当于相关滤波，被大量的快速的有竞争力的跟踪算法使用；对于核回归，无论如何，我们推导出了一种新想核相关滤波，不像其他的核算法有着和线性回归相同的复杂度；基于这个，我们还提出了快速多通道扩展的线性相关滤波，经过线性核，我们称之为双重相关滤波；这两个方法都struck或者TLD 的50个标注视频集中，排名表现突出，虽然跑在成百的pfs的视频上，并且已经被几行代码实现；为了鼓励后续开发，我们的跟踪模块已经开源；
关键词： 视觉跟踪 ，循环矩阵， 离散傅里叶变换，核方法，脊回归， 相关过滤
网友评论