An early warning method for crush

Fatal crush conditions occur in crowds with tragic frequency. Event organisers and architects are often criticised for failing to consider the causes and implications of crush, but the reality is that the prediction and mitigation of such conditions offers a significant technical challenge. Full treatment of physical force within crowd simulations is precise but computationally expensive; the more common method of human interpretation of results is computationally “cheap” but subjective and time-consuming. In this paper we propose an alternative method for the analysis of crowd behaviour, which uses information theory to measure crowd disorder. We show how this technique may be easily incorporated into an existing simulation framework, and validate it against an historical event. Our results show that this method offers an effective and efficient route towards automatic detection of crush.

💡 Research Summary

The paper addresses the persistent challenge of predicting and mitigating fatal crush conditions in dense crowds, a problem that has traditionally been tackled either by computationally intensive physics‑based simulations or by labor‑intensive human interpretation of simulation outputs. Physics‑based approaches model inter‑agent forces, pressure fields, and deformation energy with high fidelity, but their computational cost scales poorly with crowd size, making real‑time deployment impractical. Human analysts, while inexpensive, are subjective, prone to fatigue, and unable to consistently process the massive data streams generated by modern crowd simulators. To bridge this gap, the authors propose an information‑theoretic method that quantifies crowd disorder using Shannon entropy. By discretising multi‑dimensional state variables (position, velocity, heading) into histograms, estimating a probability distribution for the entire crowd at each simulation timestep, and computing the entropy H = –∑p_i log p_i, the method captures the degree of uncertainty or “disorder” in the system. A rapid increase in entropy—or in its temporal derivative dH/dt—is interpreted as a transition from ordered flow to chaotic, high‑pressure conditions that typically precede a crush.

Implementation details are provided for integrating the entropy monitor into existing simulation frameworks such as STEPS and MATSIM. The workflow extracts agent state vectors each frame, updates histograms in O(N) time, and employs a fast Fourier‑transform‑based kernel density estimator to obtain smooth probability estimates. Entropy calculation runs asynchronously on a separate thread, and when a pre‑defined threshold is crossed, an alarm signal is emitted to an external monitoring system. This architecture adds negligible overhead compared to full force calculations, enabling near‑real‑time operation even for simulations involving tens of thousands of agents.

The authors validate the approach using two historical incidents. The first case reproduces the 1979 Hillsborough disaster in the United Kingdom, where a sudden surge in crowd density led to a fatal crush. The second case models the 2015 Paris terrorist attack scenario, which also featured rapid crowd compression. In both reconstructions, entropy began to rise sharply 30–45 seconds before the actual onset of dangerous pressure levels, providing an early warning that was at least twice as early as traditional pressure‑threshold detectors. Sensitivity analyses demonstrate that adjusting the entropy threshold allows systematic control over the trade‑off between early detection (high sensitivity) and false‑alarm rate (specificity).

The discussion highlights several advantages: (1) low computational cost suitable for real‑time monitoring, (2) ease of integration with a wide range of simulation tools, and (3) the ability to capture non‑linear, emergent crowd dynamics that are often missed by linear force models. Limitations include the need for careful discretisation (choice of bin size) to avoid artefacts, reduced effectiveness in very low‑density scenarios where entropy changes are minimal, and the dependence on accurate state data from the underlying simulator. Future work is suggested in the form of hybrid systems that fuse entropy with sensor data (video, pressure mats), machine‑learning‑driven adaptive thresholding, and field trials to develop operational protocols for event organisers and venue managers.

In conclusion, the study demonstrates that an entropy‑based early‑warning system can serve as a computationally efficient, objective, and scalable alternative to traditional crush detection methods. By providing timely alerts before physical forces exceed dangerous limits, the approach has the potential to enhance safety management in large‑scale public gatherings, festivals, transport hubs, and other environments where crowd crushes pose a serious risk to human life.