Bayesian tracking by discriminant feature fusion and evolutionary importance resampling

Tracking success or failure depends heavily on how distinguishable the features separating the target object from its surroundings. From local image features, three likelihood ratio features are generated for measuring the confidence scores of pixels belonging to object or background category in this paper; and further are fused to classify them into object or background class with respect to the weights updated by chunk incremental discriminant analysis in an extended feature space adaptively. Accumulated confidence score in each candidate target region is treated as the observation likelihood for estimating object state with particle filter; meanwhile to avoid particle degeneration, an iterative importance resampling algorithm that maximizing the effective sample size by evolutionary operations is proposed to preserve particles diversity. The experimental results on tracking ground vehicles in airborne videos demonstrate the robustness of the proposed tracker on the variations of object appearance.