Particle filter tracking method based on adaptive fusion of multiple features

An algorithm for fusing multiple features adaptively in particle filter tracking framework is proposed. The tracked object is represented by a set of submodels of each feature, and then the multiple cues are combined by linear weighting on particles to obtain a more satisfying approximation at the posterior distribution of object states. According to the discriminating contribution of each feature between object and background, the confidence on each feature is adjusted, and the feature weights are estimated and updated online in order to improve the complementary between multiple features. The analyses and experiments show good performance of the proposed method against appearance and background changes under complex scenes.