Background subtraction based on adaptive non-parametric model
Qin Wan, Yaonan Wang · 2008
Object detection is an important basis for tracking and recognition in visual surveillance systems via stationary cameras. The traditional background subtraction method is difficult to detect objects accurately in the scenes, because the background is usually cluttered and not completely static. In this paper, we propose a new method for background subtraction based on adaptive non-parametric kernel density estimation. The bandwidth is chosen adaptively based on sample and estimation points, and color combing gradient are measured for pixel features. Computation complexity is also reduced by reasonable and valid assumptions. Experiments on two sequences in outdoors demonstrate that the method can model and subtract the background accurately.