Deep Embedded Clustering using Crowd Density Map

Abstract

Crowd density estimation constitutes a critical component in the effective management of smart cities. With advancements in technology, smart analytics systems are increasingly being integrated into urban monitoring processes. In this study, a model has been developed for crowd density estimation using two distinct deep learning-based methods: Congested Scene Recognition Network (CSRNet) and Deep Embedded Clustering (DEC). The performance of the proposed model has been evaluated on the ShanghaiTech Part B dataset. Initially, CSRNet was employed to generate Gaussian distributions representing crowd density from urban imagery. These distributions were constructed as structured representations of spatial crowd information and subsequently subjected to clustering analysis using DEC. The clustering performance was assessed using the silhouette score and the davies-bouldin score, yielding values of 0.54 and 0.61, respectively. Additionally, Density-Based Spatial Clustering of Applications with Noise (DBSCAN) and Gaussian Mixture Model (GMM) were applied to conduct further performance analyses. The obtained results demonstrate the efficacy of deep learning frameworks in advancing urban surveillance analytics, providing a robust foundation for the development of intelligent crowd management solutions within smart city infrastructures. © The Institution of Engineering & Technology 2024.