It has long been a big challenge to extract dense smoke regions by motion detection. As a result, there are too few suspected smoke regions being recognized for an early fire alarm. In this paper, an early smoke detecting system that can efficiently extract dense smoke regions is proposed. Firstly, since the brightness in the areas that have dynamic texture is not constant, the residuals of optical flow are calculated to locate suspected smoke regions. A certain threshold of the increment of optical flow residuals is also used to distinguish smoke from other dynamic texture. Secondly, five features that can jointly represent a smoke area, including grayish color, chrominance decrease, edge energy decrease, optical flow orientation diffusion and circularity, are chosen by thorough experiments. Experimental results show that the proposed system can detect the smoke in early time and is robust to most kinds of interferences, especially other dynamic textures.