Light-Weight Branch-Shared Multi-View Convolutional Neural Networks Crowd Counting

Crowd counting plays an important role in event planning, video surveillance, and other fields. At present, the development of single-view crowd counting is relatively mature, but due to the limitation of a single field of view, it is not suitable in some dense occluded scenes. Multi-view crowd counting uses images from multiple views to estimate the number of crowds in the current scene. Most multi-view crowd counting methods based on the deep convolutional neural networks use independent and identical view branches, which bring massive redundant features and increase the model's complexity. This paper proposes a light-weight and branch-shared convolutional neural networks method, which decreases the number of learnable parameters. This method uses the same view branch to extract multi-scale feature maps with images from different views. The camera-view feature maps will be projected into the same plane in world space to fuse, then the scene-level feature maps extracted at different scales are regressed to a scene-level density map. Extensive experiments are conducted on two public datasets (PETS2009, CityStreet), and compared with five existing methods, this method can achieve better performance.