Pedestrian counting is recently focused by many researchers in intelligent visual surveillance domain because of its wide applications. A novel method for pedestrian counting is proposed based on the foreground pixels on virtual gate. The proposed method is composed of two processes which are adaptive learning and counting. In the learning stage, many pedestrian models in the current scene are firstly extracted using a pedestrian detection method based on HOG, and then these models are used to fit a line which can be used to determine the weight of every point. In the counting stage, we get the foreground pixels on virtual gate as well as their weights, motion vectors at each frame, and the amount of pedestrian passing the virtual gate can be obtained by accumulating all those moving pixels. The experimental results show that our method has real-time performance under the premise of counting precision.