Algorithm 1: Proposed tracking algorithm.

Input: Image $I$; Initial target position $\left(u_{(0)},v_{(0)}\right)$ and scale $a^{r_0}$; previous target position $\left(u_{(t-1)},v_{(t-1)}\right)$ and scale $a^{r_{\mathrm{t}-1}}$

Output: Estimated object position $\left(u_{(t)},v_{(t)}\right)$ and scale $a^{r_t}$.

For each $I_t$

Extract the deep feature space $\widehat{x_t}$ thought the pre-trained VGG-Net;

Update matrix ${\mathbf{R}}_t$ and ${\mathbf{Q}}_t$ by linear interpolation using Equation (13) and (14). The SVD is performed and a new ${\mathbf{P}}_t$ is found;

Update the low dimensional appearance feature space ${\widehat{x}}_t$ using Equation (15);

Compute the confidence of the target position using Equation (18);

Update the tracking model ${\mathbf{A}}_t$, ${\tilde{\mathbf{b}}}_t$ and ${\widehat{x}}_t$ using Equations (19)–(22);

Compute the estimated object position $\left(u_{(t)},v_{(t)}\right)$ and scale $a^{r_t}$ using fast sub-grid detection;

If $\max \left(s_r\right)<{\theta }_{re}$,

Update the estimated object position and scale using the offline Siamese tracker;

Else

Output the estimated object position and scale directly;

End