The background subtraction is an important method to detect the moving objects, the difficulty in which is the background reconstruction. Therefore an improved background reconstruction algorithm based on pixel intensity classification is proposed. According to the hypothesis that the background pixel intensity always appears in an image sequence with maximum probability, the adjacent frames will be classified as the same or different classes of intensity based on the frame difference and frame difference accumulation, and then merging procedure is run to classify the classes, finally intensity classes with maximum appearance probability are selected as the background pixel intensity values. Simulations results show that the algorithm could affectively avoid moving mixture and well reconstruct the background of vary scene. It also has quick speed, lower store space, and strong robust.