Fast Regression Convolutional Neural Network for Visual Crowd Counting

Abstract

This paper presents an improved convolutional neural network (CNN) architecture for accurate visual crowd counting in crowd images. Comprehensive analysis on the performance and inference speed of the network model are also presented. Multi-column convolutional neural network (MCNN) for visual crowd counting through predicted density map is widely used in previous works, however this method has limitation in predicting a quality density map. Instead, the proposed network is constructed by using the powerful CNN layers, dense layers, and one regressor node with whole image-based inference. Therefore, it is less computationally intensive and inference speed can be increased. Experiments have been conducted on Mall dataset. Moreover, benchmarking on different CNN architectures have been conducted. The proposed network shows promising counting accuracy and reasonable inference speed against the existing state-of-art approaches.