Feature prediction and model updating based on filter banks

To reflect the changes of targets appearance during the tracking procedure, a model updating algorithm based on particle filter banks is proposed. Each eigenvalue of the feature histogram is assigned with a single particle filter, which compose a filter banks. According to the posterior probability model and the observation, the residual error is used to adjust the transmission radius, and the effective particle index is introduced to ensure the diversity of particles. According to the coefficients and the average residual, two criterions are established to overcome the drastic changes in the model updating. With the updating algorithm, the proposed tracker achieves better performance then the fixed model tracker in the presence of scale changes and partial occlusions.