In this paper, we propose a fast and stable human detection method using a stereo camera system called "Subtraction Stereo" with a Histograms of Oriented Gradients (HOG) feature. The proposed method performs scanning detection window using the distance of foreground regions from the subtraction image. Fast and stable human detection is achieved by the HOG feature of foreground regions with on Real AdaBoost classifier. And the experimentally demonstrated that our solution compared normal image with subtraction images. In addition, when we computed the HOG feature, we verified the size of an important cell.