For the detection of moving objects in surveillance systems, background subtraction methods are widely used. In case the background is non-stationary, modeling the background is not a simple problem. To solve the problem, many methods are proposed. In the high traffic region such as airport and subways, however, few researches have been conducted. In this paper, we classify each pixel into four different types: still background, dynamic background, and moving object, and temporary still object. And update the background according to the result. For the classification, we analyze the temporal characteristics of each pixel's intensity with likelihood test. With public video data, we experimentally show that modeling based on pixel classification improves detection accuracy in public areas which has high traffic.