Particle Filter Tracker Based on ICA Distribution Models

In order to robustly track the interested object in video sequence in the presence of cluttered background, partial occlusion and illumination variations, an improved particle-filtering tracking algorithm is proposed. Color-based tracking is limited by the lack of spatial information, which makes it difficult to discriminate between targets with similar color properties. To reduce this problem, a target model based on ICA distribution is proposed. The feature description is transformed into the feature space of ICA (independent component analysis) from color histograms. Because illuminated change is the same components by ICA transformation. The method can reduce the influence by changes in scene illumination or similar color regions, and improve the robustness of the tracker. What’s more, According to particles impoverishment phenomenon, the proposed method embeds mean-shift into the tracking frame of the particle filter algorithm. After propagated by system, each sample are herded to local max of its neighborhood，it require fewer particles to maintain multiple hypotheses and tackle the impoverished particles problem, at the same time, the computation time is efficiently reduced. It ensures performance simultaneously and accuracy. The experimental results show that the proposed method can effectively solve the tracking problem like illumination changes, partial occlusions and background scene changes.