Big Crowd Counting and Density Estimation in Still Images Using Convolutional Neural Networks

Abstract

Automatic crowd monitoring is of paramount importance in major events such as pilgrimage and large sport events where millions of people gather. This importance has become more prevalent with the current COVID-19 pandemic for better management, safety and security of the crowds. Automated crowd counting is an essential building block of crowd monitoring that aims at reducing manual counting by human operators which has become impractical. Most of existing automatic crowd counting methods either fail in high-density crowd images, ignore or partially consider spatial information, or employ heavy network architectures. In this research, we address the crowd counting problem in still images following a density estimation based approach. We employ a single-column Convolutional Neural Network (CNN) consisting of two main blocks: an encoding and a decoding block. Our model accepts the whole crowd image as input, and outputs the predicted density map from which the count is calculated. We demonstrate our approach on one of the largest and diverse datasets: ShanghaiTech dataset. Extensive experiments show the effectiveness of the proposed approach compared to recent state-of-the-art methods, being in the top four methods on all of the evaluation criteria.

Date of Award	2021
Original language	American English
Awarding Institution	HBKU College of Science and Engineering