Adaptive Weighting of Local Classifiers by Particle Filter

This paper presents adaptive weighting method for combining local classifiers by particle filter. In recent years, the effectiveness of combination of local classifiers (features) is reported. However, those methods can not cope with partial occlusion or shadows by illumination direction changes, because the stable weight is used for combining local classifiers. To be robust to them, the weight should be changed adoptively. Namely, we must select the good weight set given high likelihood from the weight space adoptively. For this purpose, particle filter is used. Each particle corresponds to the weight set for combining local classifiers. By selecting the particle (weight set) given high likelihood in current situation, the proposed method can cope with partial occlusion. The proposed method is applied to face tracking problem. Performance is evaluated by using the test sequence that the occluded area is changed dynamically. The proposed method decreases the weight for occluded region automatically, and it can track face under partial occlusion. Effectiveness of the proposed method is shown by comparison with stable weight set used in conventional methods