High-Speed Tracking with Kernelized Correlation Filters

Linear regression

$$f(z) = w^Tz$$
$$min_w\Sigma_i(f(x_i) - y_i)^2 + \lambda||w||^2$$
$$w=(X^TX+\lambda I)^{-1}X^Ty$$
$$w* = (X^HX + \lambda I)^{-1}X^Hy$$

Circulant matrices

$$X = F^Hdiag(\hat{x})F$$
$$\hat{x} = F(x)$$
$x_i$ in linear regression corresponds to a vector $x=[x_1,...,x_n]$
So, we can use circulant matrice X to represent the vector x.

Compute $\hat{w}^*$

$$X^HX = F^Hdiag(\hat{x}^*)FF^Hdiag(\hat{x})F$$
$$X^HX = F^Hdiag(\hat{x}^*\cdot\hat{x})F$$

However, this formula isn’t used in the code. The following one is used.

kernel trick

$$w = \Sigma_i\alpha_i\varphi(x_i)$$
$$\varphi^T(x)\varphi(x') = k(x,x')$$
k is the kernel function (Gaussian or Polynomial)
$$f(z) = w^T\phi(z)$$
$$w = \Sigma_i\alpha_i\phi(x_i)$$
$$f(z) = [\Sigma_i\alpha_i\phi(x_i)]\phi(z)=\Sigma_i[\alpha_i\phi(x_i)\phi(z)]$$
$$f(z) = \Sigma_i\alpha_ik(z,x_i)$$
then we need to compute $\alpha$

compute $\alpha$

$$\alpha = (K + \lambda I)^{-1}$$

compute the response map of current frame

z represent the current frame

$$K^z=C(k^{xz})$$
use circulant matrice K to represent the kernel correlation $k^{xz}$

summary

1. image patch => hog feature
2. hog feature => fft
3. kf = gaussian_correlation(xf, xf, kernel.sigma); (compute the kernel correlation of the first frame)
4. alphaf = yf ./ (kf + lambda); (compute $\alpha$
5. kzf = gaussian_correlation(zf, model_xf, kernel.sigma); (second frame)
   response = real(ifft2(model_alphaf .* kzf));
   as illustrated in the formula