Crowd detection and counting from images using MResnet

Abstract

Crowd Detection and counting is important for crowd control and monitoring in places like pilgrimages. Automatic crowd detection from images have several challenges. Different scale variations and viewpoints of images make it difficult for crowd detection models to generalize for broader data. Most of the existing approaches for crowd detection contains multiple columns for extracting multi-scale features. By using multiple columns through a deeper network can cause the layers to lose features as the layers get deeper. In this paper, a new Multi-Residual Network (MResnet) is proposed for crowd detection and counting. MResnet contains multiple three columns sub-networks with three receptive field variations. The advantages of the proposed network is that each sub-network has a specific receptive field for imbalanced distribution of human crowd in the image. Residual connections are utilized in each subnetwork for information propagation. The MResnet is evaluated using the ShanghaiTech dataset. Extensive experiments have shown that our proposed network achieves lower count error and high spatial localization.