Artificial Intelligence Guided Crowd Counting and Density Estimation with Enhanced CSRNet

Abstract

Estimating crowd size in dense environments re-mains a complex problem, yet it holds critical value for safety monitoring, city infrastructure design, and large-scale gathering coordination. Leveraging contemporary developments in neural network architectures and machine intelligence, researchers have markedly enhanced the precision of population counts derived from both still imagery and motion footage. This research focuses on the development of an enhanced deep neural network-based crowd counting model. Since the original CSRNet primarily relies on head detection for crowd estimation, an additional face detection module has been incorporated to improve its capability in scenarios where facial features are visible. The proposed enhancement increases the flexibility and estimation accuracy of CSRNet by integrating individual face detection with crowd density estimation. Furthermore, this study presents a comprehensive comparative analysis of the proposed enhanced CSRNet against three state-of-the-art crowd counting models, namely Bayesian Network (BAYNet), Distribution Matching (DM-Count), and Scale Aggregation Feature Attention Network (SFANet). The models are evaluated on multiple datasets to investigate their accuracy, robustness, adaptability to varying crowd densities, and performance under complex environmental conditions. The comparison also highlights the methodological differences and computational characteristics of these approaches. Model performance is assessed using the standard evaluation metrics of Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). Experimental results demonstrate that the proposed enhanced CSRNet achieves a competitive RMSE/MAE of 9.61/93.05 on 64 × 64 images of the custom dataset, outperforming the baseline models and demonstrating its effectiveness for accurate crowd counting.