Statistical Background Subtraction Based on the Exact Per-pixel Distributions.
Youngbae Hwang, Hanbyul Joo, Jun‐Sik Kim, In-So Kweon · 2007
Most of background subtraction methods represent background statistics using probabilistic unified frameworks such as the Gaussian mixture model or kernel density estimation. But these models cannot define the exact difference between two pixels. It causes misclassification such as false alarms and misses. We presented a new sensor noise model appropriate for general CCD cameras. Based on this, we propose a novel background subtraction method. Our noise modeling needs a line estimation step to relate image intensities with parameters of the noise distribution. This paper describes a new line estimation algorithm given two consecutive static images, and from which can have a well-fitted distribution for each pixel according to intensity of the pixel. In addition, we present a background update method to deal with the continuous variation of the background. We can estimate accurate foregrounds by adapting the estimated per-pixel distributions and background updates. 1.