Authors:
S. Rubin Bose, J. Angelin Jeba, O. Jeba Singh, R. Regin, S. Suman Rajest, G. Mary Amirtha Sagayee
Addresses:
School of Computer Science and Engineering, SRM Institute of Science and Technology, Ramapuram, Chennai, Tamil Nadu, India. Department of Electronics and Communication Engineering, S.A. Engineering College, Chennai, Tamil Nadu, India. Centre for Academic Research, Alliance University, Bengaluru, Karnataka, India. Department of Research and Development, Dhaanish Ahmed College of Engineering, Chennai, Tamil Nadu, India. Department of Computing and Information Sciences, University of Technology and Applied Sciences, Muscat, Oman.
Abstract:
Stampedes in densely populated gatherings such as festivals, stadiums, and religious events remain a major safety concern, often resulting in severe casualties and chaotic crowd behavior. Traditional surveillance and rule-based crowd control systems struggle to identify potential hazards early because they cannot capture dynamic spatial–temporal patterns. To overcome these limitations, this research proposes a machine-learning-based predictive framework that anticipates stampede-prone conditions by continuously analyzing crowd density, velocity, and movement direction. The system employs a hybrid CNN–LSTM architecture, where the CNN module extracts spatial features from density and optical flow maps, while the LSTM component models temporal dependencies to detect evolving risk trends. This architecture enables the model to understand both localized crowd concentration and collective movement behavior across time, offering proactive alerts before congestion escalates. The proposed model was trained and evaluated using large-scale crowd datasets and simulated event footage, achieving an overall accuracy of 98.7% and a 28% reduction in false alarms compared to existing machine learning approaches such as Random Forest and Support Vector Regression. These results demonstrate the superior precision, robustness, and scalability of the proposed system for real-time crowd-safety monitoring. The main challenges encountered during development involved managing data imbalance between safe and risky instances, optimizing computation for real-time processing, and ensuring consistent accuracy under varying lighting and environmental conditions.
Keywords: Crowd Safety; Machine Learning; Crowd Safety Monitoring; Convolutional Neural Network (CNN); Crowd Density; Optical Flow; Environmental Conditions; Random Forest; Long Short-Term Memory (LSTM).
Received on: 28/12/2024, Revised on: 17/02/2025, Accepted on: 10/05/2025, Published on: 03/01/2026
DOI: 10.64091/ATICL.2026.000255
AVE Trends in Intelligent Computer Letters, 2026 Vol. 2 No. 1 , Pages: 53–66