Input: image sequence, particle set |
Output: particle set and the object state |
Initialization: |
(1) Initializing the tracked object manually. |
(2) Constructing the particle set . |
(3) Calculating the template . |
Object tracking: |
(1) for from 1 to the last frame do |
(2) Estimate the object motion utilizing the proposed insect vision inspired detector and |
obtain the response . |
(3) Local search strategy: |
(4) Propagating: According to the proposed state transition model (7), |
propagate each particle state to get a new state for time . |
(5) Re-sampling: Based on the state and the template, compute the weight of each |
propagated particle at time , and then normalize the particles weight. |
Generate a new particle set . |
(6) Lost detection: (to address severe occlusion and drifts problems) |
(7) Count the number of valid samples and test the Two-Prerequisites. If the Two-Prerequisites are satisfied, |
let LF = 1, otherwise, LF = 0. (LF is the label to mark the lost.) |
(8) Global search strategy: |
(9) if LF = 1 (the tracker loses the object) or for every frames then |
(10) Refining: Search the interest image region where to seek for a more accurate state . |
(11) Weighting: Based on the state and the template, compute the weight of each particle at time |
and normalize it. |
(12) end if |
(13) Estimating: |
(14) Calculate the object position at time and get the best state: , where . |
(15) Online template update: |
(16) Update the template according to (15) for every frames. |
(17) end for |