The task of crowd counting is to automatically estimate the pedestrian numberin crowd images. To cope with the scale and perspective changes that commonlyexist in crowd images, state-of-the-art approaches employ multi-column CNNarchitectures to regress density maps of crowd images. Multiple columns havedifferent receptive fields corresponding to pedestrians (heads) of differentscales. We instead propose a scale-adaptive CNN (SaCNN) architecture with abackbone of fixed small receptive fields. We extract feature maps from multiplelayers and adapt them to have the same output size; we combine them to producethe final density map. The number of people is computed by integrating thedensity map. We also introduce a relative count loss along with the density maploss to improve the network generalization on crowd scenes with fewpedestrians, where most representative approaches perform poorly on. We conductextensive experiments on the ShanghaiTech, UCF_CC_50 and WorldExpo datasets aswell as a new dataset SmartCity that we collect for crowd scenes with fewpeople. The results demonstrate significant improvements of SaCNN over thestate-of-the-art.
展开▼
