Skip to main content
. 2016 Sep 7;16(9):1443. doi: 10.3390/s16091443
Algorithm 2. The Blur KCF tracker
  • Inputs: 

    Template x, target position x0 initialized by the given ground truth in the first frame;

    Initial scale s = 1;

    Corresponding image sequence for tracking;

  • Outputs: 

    Estimated target position x0′ and estimated target scale s′ for each frame;

    Updated x′ and model coefficient α′ at each time;

  • 1:

    Selection: Select kernel function and features type;

  • 2:

    Initialize the model coefficient α with Equation (5);

  • 3:

    repeat;

  • 4:

    Calculate the kernel function Kxx′ with Equation (6) or (7) according to the selected kernel;

  • 5:

    Calculate the response f(z) with Equation (8) and acquire the maximum response Rmax;

  • 6:

    Compute the Clarity value c of the candidate image patch via JNB metric;

  • 7:

    Find the threshold cτ with Equation (17); //Larger range search;

  • 8:

    If c ≤ cτ then;

  • 9:

    Make Radon Transformation for patch to figure out the target movement with Equation (15);

  • 10:

    Enlarge the search range with a shift from x0 with Equation (18) of four directions;

  • 11:

    for every new candidate image patch;

  • 12:

    Calculate the kernel function Kxx′;

  • 13:

    Calculate the response f′(z) and acquire the maximum response R′;

  • 14:

    Weighted the response with Equation (19);

  • 15:

    If R′ > Rmax then Rmax = R′;

  • 16:

    end;

  • 17:

    end if;

  • 18:

    Acquire the position of the maximum response x0′ and new size s′ with Equations (10) and (11);

  • 19:

    Update the template by Equations (5) and (9);

  • 20:

    Until End of video sequences;

  • 21:

    Return {x′, α′, s′}.