"Computer Vision" particle filter tracking

Particle filtering steps

1. Initialize randomly select N points, weights uniform assignment 1/n

2. Select the target feature, color histogram, etc., to obtain a priori probability density, compare the similarity degree

3. Determine the state transition matrix for predicting the next frame target location

Cycle start

4. According to the state transfer matrix, for each particle, predict the target new position

5. Obtain the system observations and calculate the characteristics at the observation location

6. Calculate the characteristics of each prediction location, compared with the position characteristics of the observation place, the higher the similarity degree assigns the larger weight, conversely, gives the smaller weight

7. Get the best position of the current target for particle weighting

8. Particle resampling

End of cycle

Above is my own understanding, some questions still do not understand:

1. Have the observed value also corrected, the result of the correction is really better?

2. Importance sampling in particle filtering theory is completely out of the question, where does the importance sampling work?

3. In particle filter theory, why do we want to simulate the posteriori probability, what is the actual meaning? Where does the tracking process manifest itself?

"Computer Vision" particle filter tracking