A conservative scene model update policy

In this paper, we present a new pixel-level scene model for segmenting video into foreground and background structure. The design of the model is largely influenced by several recently reported stochastic background models that have been shown to significantly outperform traditional deterministic techniques. In contrast to existing nonparametric scene models, we propose a learning algorithm that integrates new information into the models by replacing the most significant outlying values with respect to the current sample collections. Outliers are identified using a variable bandwidth kernel density estimation (KDE) procedure. We demonstrate the superiority of our model against a recent state-of-the-art video segmentation system and compare and contrast the theoretical aspects of our model with a wide variety of existing techniques, and well known video segmentation challenges.