Abstract: This paper presents a novel approach to the task of video-based crowd counting, which can be formalized as the regression problem of learning a mapping from an input image to an output crowd density map. Convolutional neural networks (CNNs) have demonstrated striking accuracy gains in a range of computer vision tasks, including crowd counting. However, the dominant focus within the crowd counting literature has been on the single-frame case or applying CNNs to videos in a frame-by-frame fashion without leveraging motion information. This paper proposes a novel architecture that exploits the spatiotemporal information captured in a video stream by combining an optical flow pyramid with an appearance-based CNN. Extensive empirical evaluation on five public datasets comparing against numerous state-of-the-art approaches demonstrates the efficacy of the proposed architecture, with our methods reporting best results on all datasets. Finally, a set of transfer learning experiments shows that, once the proposed model is trained on one dataset, it can be transferred to another using a limited number of training examples and still exhibit high accuracy.