Robust model adaptation for tracking with online weighted color and shape feature

Histogram-based particle filters have emerged as an appealing method for the target tracking. As colour and shape features are widely used to represent the target, we propose in this paper a novel method to combine these two features by assigning an adaptive weighting factor to each feature, in a particle filtering framework. In other words, the feature with higher likelihood and the property of high saliency will contribute more than other features to estimate the posterior density function of the target state in tracking. To cope with the target appearance changes, our tracker extracts the contextual information from the background to alleviate model drifting problem. The contextual information is therefore used for reference model update. We tested our proposed algorithm on some publicly available datasets, and the results from these video sequences have shown that the proposed tracker can tackle several open problems in tracking including heavy illumination changes, dramatic self-deformation and background clutter.