Multifeatures Based Compressive Sensing Tracking

Algorithm 1

Multifeature based compressive sensing tracking.

Input:

 (i) Current visible frame 𝐼^𝑘_𝜐∈𝑅^ℎ×𝑤 and infrared frame 𝐼^𝑘_ir∈𝑅^ℎ×𝑤

 (ii) Particles 𝑠^𝑖_𝑘−1,𝑖=1,2,…,𝑁_𝑠

 (iii) Particles Measurement matrix Φ∈𝑅^𝑚×𝑛

 (iv) Template set 𝐴=[𝑇,𝐼]∈𝑅^{𝑛×(𝑛+𝑁𝑡)}

 (v) Preset parameter 𝜆>0

(1) Normalize each column of Φ𝐴;

(2) Generate particles 𝑠^𝑖_𝑘,𝑖=1,2,…,𝑁_𝑠 according to state transition distribution;

(3) for  𝑖=1→𝑁_𝑠  do

(4)    Obtain target candidates 𝐠_𝑖 and relevant candidate center 𝑦_𝑖 according to 𝐬^𝑖_𝑘;

(5)    Calculate RGB histogram 𝐩_RGB, LBP 𝐩_LBP of candidate in 𝐼^𝑘_𝜐 and intensity histogram 𝐩_gray of candidate in 𝐼^𝑘_ir;

(6)    Stack 𝐩_LBP,𝐩_gray and 𝐩_RGB into an observation 𝐳_𝑖=[𝐩_LBP,𝐩_gray,𝐩_RGB]^𝑇;

(7)    Get measurements 𝐳^′=Φ𝐳_𝑖;

(8)    Solving (17) to get coefficient vector 𝐱 ;

(9)    Calculate residual 𝐫_𝑖=Φ𝐳_𝑖−Φ𝐴𝐱_𝑡2;

(10) Obtain observation likelihood 𝑙_𝑖=exp−𝜆𝐫_𝑖;

(11) end

(12) Resample the particles 𝐬^𝑖_𝑘 according to 𝑙_𝑖;

(13) Using Maximum A Posterior (MAP) or Mean Square Error (MSE) to

  estimate the state 𝐬_𝑘 of current frame, which characterizes the current tracking target 𝐠 ;

(14) Get target observation 𝐳̂ corresponding to current tracking result 𝐠 ;

(15) Recalculate 𝐱 for 𝐳̂ by solving (17);

(16) Update target template 𝑇;

Output:

 (i) Tracked target 𝐠

 (ii) Updated target dynamic state 𝐬_𝑘

 (iii) Updated target template set 𝑇