Estimating the Size of Crowds through Deep Learning

Automatically counting the number of people in a crowd from a photographic image must address numerous potential issues: partial occlusion, differences in illumination, varying poses, and potential blurring caused by the different distances from individuals in the crowd to the camera. This is especially true with large crowds where close-up counting of faces is not easy/possible. Deep learning offers some opportunities to tackle these issues. In this paper, we describe the implementation of a crowd counting application based on a Multi-task Convolutional Neural Network. This approach enables the model to detect and gradually refine the identification and counting of faces. We tailor the model to make it suitable for the counting of individuals in larger-scale crowds using non-trivial and hence realistic and challenging crowd images. We provide details of the practical realization of the model through two mobile applications (iPhone/Android) and a web application for statistical analysis.