Adaptive Foreground Detection Based on Weighted Kernel Density Estimation

In order to avoid the impacts of moving foreground on background modeling in training stage, an adaptive foreground detection method based on weighted kernel density estimation (KDE) was proposed. In this method, temporal stable pixels are assigned more weights, and a weighted KDE background model is established to reduce the interference of foreground during background model building. Based on this background model, a strategy for dynamic foreground threshold was proposed. With the spatial consistency of foreground, "holes" in foreground are filled and thresholds are updated in the same time. The experimental results show that the proposed foreground detection method is able to achieve over 90% precise and recall rates in various scenes even under the condition that there are moving objects, and it outperforms the conventional background subtraction methods.