A model for dynamic object segmentation with kernel density estimation based on gradient features

The dynamic object segmentation in videos taken from a static camera is a basic technique in many vision surveillance applications. In order to suppress fake objects caused by dynamic cast shadows and reflection images, this paper presents a novel segmentation model with the function of cast shadow and reflection image suppression. This model is a kernel density estimation model based on dynamic gradient features. Unlike the conventional kernel density estimation model which can only suppress cast shadows in color videos, this model can also suppress them in intensity videos, and under the circumstance of diffusion it can suppress reflection images effectively. Although this model may cause the increase of the false negative rate, its function of fake object suppression is remarkable. Furthermore, the false negative rate can be reduced with other convenient methods. Some experimental results by real videos are also presented in this paper to demonstrate the effectiveness of this model.