Part-Based Visual Tracking via Online Weighted P-N Learning

Algorithm 1

Tracking based on proposed method.

  Initialization:

(1) Segment object into 𝑁 blocks and obtain 𝐶𝐵={𝑐𝑏_𝑖}^𝑁_𝑖=1;

(2) Select the set of LFBs 𝑆𝐵={𝑠𝑏_𝑘}^𝑀_𝑘=1(𝑀<𝑁);

(3) Compute the set of offset in first frame {𝑧¹_𝑘}^𝑀_𝑘=1;

(4) Generate positive samples (𝒳⁺_𝑘,𝒴⁺_𝑘) and negative

  samples   (𝒳⁻_𝑘,𝒴⁻_𝑘) for the 𝑘^𝑡ℎ LFB;

(5) Train classifier 𝑓_𝑘 for the 𝑘^th LFB with data (𝑋,𝑌) via

   weighted P-N learning, where 𝑋={𝒳⁺_𝑘,𝒳⁻_𝑘},

   and 𝑌={𝒴⁺_𝑘,𝒴⁻_𝑘};

  Object Tracking:

(6) for   𝑡=2 to the end of the sequence  do

(7) for   𝑘=1 to 𝑀  do

(8)   Estimate 𝑠𝑏^𝑡_𝑘 via detecting 𝑘^th LFB with classifier

     in LK framework;

(9)   Compute 𝐵^𝑡_𝑘 via (15);

(10)  Retrain classifier 𝑓_𝑖;

(11) end for

(12) Estimate 𝐵^𝑡 via (16);

(13) Adjust the set of offset via (14);

(14) for   𝑘=1 to 𝑀  do

(15)  Check each LFB with corresponding classifier

(16)  if   𝑃^𝑘_avg<𝜆  do

(17)   Update 𝑠𝑏_𝑘←𝑢𝑏_𝜏 based on Section 4.4;

(18)   Generate positive samples (𝒳⁺_𝑘,𝒴⁺_𝑘) and

       negative  samples (𝒳⁻_𝑘,𝒴⁻_𝑘) for the 𝑘^th LFB;

(19)    Train new classifier 𝑓_𝑘 for the 𝑘^th LFB

(20)    break;

(21)   end if

(22) end for

   End