(1) Extract target regions in the first frames
 (1.1) Perform background subtraction and get motion region.
 (1.2) Building a pedestrian detector using HOG descriptors and SVM.
 (1.3) Detect pedestrians in each frame.
(2) Build up the discriminative appearance model
 For
  (2.1) Randomly sample patches
   Generate superpixel patches and LBP patches for each target.
   Extract superpixel descriptor and LBP descriptor for all patches.
 End For
 For target
  (2.2) Construct Codebook
   perform meanshift clustering for all superpixel descriptor, cluster centers
   compose the superpixel codebook.
   perform meanshift clustering for all LBP descriptor, cluster centers compose
    the LBP codebook.
   For
    Find its nearest keyword and make statistics of the appearance times of the keyword.
   End For
   Compose trained bags .
 End For
(3) Tracking
 For target
  (3.1) Initialize particle state distribution using the center of specifying region.
  (3.2) Set initial weight value of feature information a = b = 0.
 End For
 For
  For target
   (3.3) Important sampling step
    Propagate and get new particles using (7).
   (3.4) Update the weights
    Compute the observation likelihood function and for each
    particle using (11).
    Update weights value of features information using (12).
    If
     
    End if
  End For
 End For
 (3.5) Update codebook
  For each frames
   Perform mean shift clustering again on and the old codebook using (3).
  End For
 (3.6) State estimation
  Estimate the state
Algorithm 1: Our algorithm.