Supplementary material : Particle dynamics and multi - channel feature dictionaries for robust visual tracking

The basic idea of KLD-sampling [3] is to find the number of particles in each iteration such that the error between the true posterior probability density and the probability density approximated by the particle filter is less than ν with probability (1−δ ). At any particular iteration, suppose we draw n particles from a discrete probability distribution that has k disparate bins. Defining the vector N = [N1,N2, . . . ,Nk] as the number of particles drawn from each bin, we can see that N follows a multinomial distribution fk(n,p), where p = [p1, p2, . . . , pk] represents the probability of each of the k bins. We can use the maximum