Deep Learning for Scene-Independent Crowd Analysis

In large cities with high population densities, the assembly of large crowds in public events and public areas increases risks to public safety and transportation, which have become major concerns to the community. Although much effort has been made on crowd scene understanding, many of existing works are scene-specific, i.e., models learned from a particular scene cannot be well applied to other scenes. It limits the application of these technologies, since extra training samples have to be collected from a new scene. This book chapter will introduce scene-independent crowd analysis with deep learning. Once generic deep models are learned from large-scale training sets, they can be applied to various crowd scenes without being trained again. The topics cover crowd density estimation, crowd counting, and crowd attribute recognition. Deep learning is driven by large-scale training. Several large-scale datasets developed recently are introduced. Multiple issues related to building crowd datasets, such as annotation and increasing scene diversity, will be discussed. We will also introduce multiple architectures of deep neural networks and training strategies to learn the feature representations for crowd analysis.