Video crowd counting is to estimate the number of people at a crowded scene in a video sequence. We design a network that achieves high counting accuracy and inference efficiency by exploiting people flow information (the movement of people in the video). Previous approaches on crowd counting often have not accounted for crowd flow sufficiently, or suffer from inference overhead or flow-dependent parameter tuning. We propose Flow-Count, a novel multi-task network based on density map regression that properly correlates flow features with crowd count. Flow-Count learns a pixel-level people flow map as an explicit auxiliary supervision signal to effectively capture people flow in detail, while such flow estimation is not needed in the inference stage. Extensive experiments conducted on re...[ Read more ]

Video crowd counting is to estimate the number of people at a crowded scene in a video sequence. We design a network that achieves high counting accuracy and inference efficiency by exploiting people flow information (the movement of people in the video). Previous approaches on crowd counting often have not accounted for crowd flow sufficiently, or suffer from inference overhead or flow-dependent parameter tuning. We propose Flow-Count, a novel multi-task network based on density map regression that properly correlates flow features with crowd count. Flow-Count learns a pixel-level people flow map as an explicit auxiliary supervision signal to effectively capture people flow in detail, while such flow estimation is not needed in the inference stage. Extensive experiments conducted on representative video datasets demonstrate that Flow-Count, as compared with state-of-the-art schemes, greatly reduces crowd counting errors by 18.66% for CroHD dataset and 9.5% for VSCrowd dataset respectively, while being 21.5% faster at inference stage.