The basic steps for computer vision-based automatic video analysis are to detect and track objects. In order to do these steps, the most important and commonly used methods are background subtraction methods. This paper proposes a novel background subtraction method, which is a member of estimation-based background model, involving robust regression technique. The method proposed can estimate backgrounds at enough precision even when there are foreground objects stationary for a long time, which is often the case in images belonging to urban traffic cameras. The method has been tested with existing datasets in the literature and proved its success compared with other known methods. Moreover, it has been tested also with the dataset prepared during this research, which involves images where vehicles stop in different periods and then move again.