Abstract: The Histogram-of-Oriented-Gradient (HOG) is a widely used feature used in many pattern recognition applications involving pedestrian detection. The basic idea of HOG is that the local pedestrian appearance can be characterized by the distribution of local intensity gradients and edge directions. In this letter we describe a dual superpixel HOG algorithm in which we fuse together two HOG feature vectors. The first vector is the traditional HOG feature vector calculated on the input image. The second vector is a HOG feature vector which is calculated on the input image after superpixel segmentation. By fusing the two HOG vectors together we obtain a fused HOG with an enhanced performance while at the same time being fully compatible with the traditional HOG. Experimental results on standard pedestrian detection databases show that for noisy input the dual HOG significantly outperforms the traditional HOG detector. Keywords: Histogram-of-oriented-gradient, Pedestrian detection, superpixel, HOG, image processing