Summary of High-Speed Tracking with Kernelized Correlation Filters

Summary of

High-Speed Tracking with Kernelized Correlation Filters

João F. Henriques, Rui Caseiro, Pedro Martins, and Jorge Batista

Summarize by Wenju Huang

Abstract

With the theory of circulant matrices, this paper present a classification trained with all patches form a single image, which can reduce the redundancy of the random sample in the normal classification. What’s more, the characteristic is mapped to high-dimensional space using the kernel trick. According to the kinds of the kernel function and the dimensionality of input futures, we get the different filters, such as MOSSE(linear kernel, one-dimensional future), KCF(non-linear kernel), and DCK(linear kernel, high-dimensional future).

1. Pre-knowledge

1) Regularized Least-Squares Classification (Ridge Regression)[1]

Compared with the ordinary least squares regression, the RLSC considers the sensitivity of the regression parameters w. The goal of training is to find a function f(z) = w^Tz that minimizes the squared error over samples x_i and their regression targets y_i.

 

the lamda is a trades off parameter between the the empirical risk and ||W||².

2) The theory of Circulant matrices[2]

In linear algebra, a circulant matrix is a special kind of Toeplitz matrix where each row vector is rotated one element to the right relative to the preceding row vector.

An n*n circulant matrices C takes the form

 

Given a matrix equation

Cx = b

where C is a circulant square matrix of size n×n we can write the equation as the circular convolution

c*x = b

where c is the first column of C, and the vectors c, x and b are cyclically extended in each direction.

3) The kernel trick

Features which is linearly non-separable in the low-dimension space, will be  linearly separated when map them to the high-dimension space. However, When operation directly in the high-dimension space, we will encounter the curse of dimensionality problem. Fortunately, the kernel trick can solve this problem. According the kernel trick, we have:

K(x,z) =<Φ(x),Φ(z) >

The above function means that we can use a suitable kernel function to get the product of two features in the the high-dimension space and don’t have to know the really high-dimension space. Note that different kernel function relative to the different high-dimension space, so we have to choose the kernel function according our feature need. Here some common kernel function.

Epanechnikov Kernel:

    

 

Uniform Kernel:

               

Normal Kernel:

                

2. Major formulation and their inference

The finally output function is:

 

where f(z) is a vector, containing the output for all cyclic shifts of z., k^xz is the kernel values and α is a learned coefficient.

k^xz  take the form:

Gaussian kernels

 

     Polynomial kernels

   linear kernel

α can be obtained by

 

where lambda is a regularization parameter, and k^xx is the kernel values like we introduce above. In fact, there is only the above three equations in our final algorithm. lambda is figured form the Regularized Least-Squares Classification with the circulant matrices and kernel trick.

 

[1]R. Rifkin, G. Yeo, and T. Poggio, “Regularized least-squares classification,”Nato Science Series Sub Series III Computer and Systems Sciences, vol. 190, pp. 131–154, 2003.

[2]wiki https://en.wikipedia.org/wiki/Circulant_matrix