We used a parametric representation of the object of interest
based on its color appearance due to its discrimination power and
also because it provides a reduced computational cost. The
representation is then convolved with a Gaussian filter to eliminate bad information on the edges. The tracking problem starts with an initialization process of the population in a reduced search area around the object's previous location. In order to enhance the quality of the population, we integrated two special candidates:the previous location of the object and the predicted solution
using the prediction step of the Kalman Filter. Each solution of the
population represents a candidate region of the current image.
Regions are compared to the object of interest in terms of the
parametric representation using the Bhattacharyya distance and a
Lévy flight is then performed around the current best solution to
generate a new candidate. This new element is compared to a
randomly chosen solution from the population. The discovery
process described in the original algorithm is also included to