Algorithm 1. Aojmus.

Input:

Current frame It; The video for tracking Svideos;

Target position $p_{t-1}$ and scale $s_{t-1}$ of previous frame;

The target feature template, ${\hat{x}}_{t-1}$ and the filter template, ${\hat{\alpha }}_{t-1}$;

The scale model $A_{t-1}^{scale}$, $B_{t-1}^{scale}$.

Output:

Target position pt and scale st of current frame;

The updated target feature template, ${\hat{x}}_t$ and filter template, ${\hat{\alpha }}_t$;

The updated scale model $A_t^{scale}$, $B_t^{scale}$.

1:  for each $I_t\in S_{videos}$ do

2:    Sample the new patch $z^t$ from It at $p_{t-1}$;

3:    Extract a scale sample zscale from It at pt and $s_{t-1}$;

4:    Extract the HOG and CN features and fused with Eq. (7);

5:    Calculate the response $\hat{f}\left(z^t\right)$ with Eq. (5), and get Fmax;

6:    Calculate $N$, WAPCE and RSFM with Eqs. (12)–(14), and adaptively judge whether there is occlusion and the scope;

7:    Get the ${\xi }_R$ and ξ by Eqs. (16) and (17);

8:    Compute the scale correlations yscale using zscale, $A_{t-1}^{scale}$ and $B_{t-1}^{scale}$ in Eq. (10);

9:    Set st to the maximum of yscale;

10:   Use Eq. (18) to update ${\hat{x}}_t$ and ${\hat{\alpha }}_t$ with ${\hat{x}}_{t-1}$ and ${\hat{\alpha }}_{t-1}$ adaptively;

11:   Use Eq. (9) to update $A_t^{scale}$ and $B_t^{scale}$ with $A_{t-1}^{scale}$ and $B_{t-1}^{scale}$.

12:   Return pt and the updated ${\hat{x}}_t$, ${\hat{\alpha }}_t$, $A_t^{scale}$, $B_t^{scale}$.

13:  end for