A Sampling-Resampling Based Bayesian Learning Approach for Object Tracking

This paper proposes an effective background subtraction technique in still camera videos, to track objects with high degree of sensitivity, accuracy and low false detections. The method involves applying a Bayesian learning technique to update parameters of clusters formed by pixel observations at a particular spatial position. The proposed method also overcomes the limitation of having a heuristically fixed number of clusters in existing tracking techniques which are based on mixture modeling of background. The results favourably compare with some existing methods for a variety of test videos, including those having very low object-background contrast.