Analysis of Crowd using CNN with Physical distance Status

Abstract

Crowd analysis plays a critical role in public safety, resource allocation, and effective crowd management during large gatherings. This paper presents an enhanced convolutional neural network (CNN)-based model for crowd estimation that incorporates physical distance status, a key parameter for post-pandemic safety requirements. The proposed approach performs crowd density classification and individual count estimation, categorizing crowd levels into four classes based on count and proximity. The model is evaluated on the NWPU-Crowd dataset using standard performance metrics including accuracy, Mean Absolute Error (MAE), and Root Mean Square Error (RMSE). It achieves a significant improvement over existing methods, with an MAE of 3.12 and RMSE of 4.28, and precision above 95% across all classes. A comprehensive study confirms the benefit of integrating physical distance features. The system's robustness is further demonstrated through visual analysis and comparative performance against state-of-the-art models. This research contributes toward intelligent, safety-aware crowd monitoring systems suitable for real-time deployment using edge-based AI.