vol 112(3) / pp 231-242

We present a new noise model for color channels for statistical change detection. Based on this noise
modeling, we estimate the distribution of Euclidean distances between the pixel colors of the background
image and those of the foreground image. The optimal threshold for change detection is automatically
determined using the estimated distribution. We show that our noise modeling is appropriate for various
color spaces. Because the detection results differ according to the color space, we utilize the expected
number of error pixels to select the appropriate color space for our method. Even if we detect changes
based on the optimal threshold in a properly selected color space, there will inevitably be some false classifications.
To reject these erroneous cases, we adopt graph cuts that efficiently minimize the global
energy while taking into account the effect of neighboring pixels. To validate the proposed method, we
show experimental results for a large number of images including indoor and outdoor scenes with complex
clutter.