Efficient particle filter using non-stationary Gaussian based model

In this paper, efficient model estimation based on particle filter for visual object tracking in automated video surveillance is proposed. Particle filter is used for our object prediction. We try to reduce complexity of particle filter by embedding non-stationary Gaussian object model into particle filter algorithm in that the dimension of the object state can be reduced. The object model is based on adaptive bounding box feature that can be used to handle non-rigid object tracking. The bounding box feature is defined as follows: image coordinate (x, y), recursive width (w) and recursive height (h). The recursive width and height will be updated by using non-stationary Gaussian formula. We use the recursive width and height as the fixed value for a set of the samples in the particle filter process. This can reduce the number of the samples improving complexity of the particle filter process. The proposed particle filter is compared with color-based particle filter to validate the proposed object model. From the experimental results, we can get the better accuracy while the number of the samples is maintained.