Crowd simulation technology
In critical infrastructures such as railway and metro infrastructures, the design, development, and testing of safety- and security-related resilience and mitigation strategies need to be exhaustively accurate. This often requires time-consuming and costly physical trial runs to effectively evaluate and refine the developed technologies.
Crowd simulation provides a testbed for the development of security and safety measures and mitigation strategies by providing time-scaled simulations to visualize the fallout and disruptions caused by various attack scenarios, and the outcome of applying the studied strategies. A simulation platform, such as the iCrowd simulator, can be used to provide an overview of how the scenario unravels faster than real-time, observe the effect of real-time adjustments to the simulation and determine chain reactions. Eventually this aims to reduce the cost of developing and testing existing infrastructure and processes as well as proposed changes on them, and marginally improve disaster response by reducing or avoiding fatalities in the railway sector.
Crowd simulation was used in the context of SAFETY4RAILS to detect possible vulnerabilities in multi-modal railway and metro stations, extract useful metrics for their evaluation, and by doing so provide a risk assessment tool to develop better resilience strategies for the safety and security of the users of such infrastructures. This is achieved by simulating the impact of cyber-physical threats on metro and railway stations, taking into account the crowd’s behavior and external interconnected infrastructures, simulating realistic information propagation models, and using the infrastructure’s current surveillance and security policies, such as evacuation processes, CCTV systems, security personnel positioning, etc.
iCrowd in SAFETY4RAILS
The iCrowd Simulator of the National Center for Scientific Research «Demokritos», initiallydesigned and developed in the context of the TASS Project  for simulating an emergency crowd evacuation at an airport  is a general purpose Agent-Based modeling platform  aiming to provide an abstract, domain-agnostic simulation framework. It implements a modern, multithreaded, data-oriented simulation engine employing the latest state-of-the-art programming technologies and paradigms. Thanks to its high performance parallel execution model, iCrowd is capable of handling very large-scale crowds. It can be utilized in any area such as building interiors and exteriors, stadiums, public places like squares, open-air festivals, or even a small city . It can be deployed on a single system with consumer-grade hardware for simple simulations, or on more advanced clusters to facilitate more complex scenarios in real time. iCrowd’s main functionalities include advanced intelligence modeling using Behavior-Trees, high-precision autonomous movement based on the ClearPath model, autonomous collision avoidance based on Velocity Obstacles and Social Forces, along with extended connectivity support and an embedded Lua interpreter for scenario scripting capabilities.
Overall, the functionality that is provided by iCrowd is the following:
- Determine infrastructure’s ability to handle high congestion levels: Simulate large crowds entering and exiting the station in normal circumstances and in case of emergency or schedule disruption. Eventually, provide a visualization of the overall movement patterns and congestions levels, all within the graphical user interface of the simulator.
- Study schedule disruptions’ impact on agents’ flow rate: Deactivate the arrivals and departures of trains in a simulation and inspect how the crowd moves around the station during high congestion levels.
- Detect and reduce crowd bottlenecks: Simulate large crowds moving in the same direction or towards the same target and visualize the congestion and pressure levels using heatmaps, in order to better understand the weak points of the station and better set up the available space.
- Validate and improve CCTV systems: Simulate cameras and guards moving around the environment and provide heatmaps and statistics regarding the detection of malicious actors.
- Validate anomaly detection systems and algorithms: Simulate malicious actors by implementing real-time dynamic detection evasion and examine the resulting movement trajectories. Provide data to eventually improve the coverage of the available space whilst minimizing the required resources .
- Validate and improve evacuation processes: Execute simulations based on real measurements to establish a baseline, and then run more simulations with adjusted parameters to examine alternate evacuation processes in conjunction with other event simulators, e.g., fire and injury simulators , bomb blast simulators, etc.
- Implement and measure user-defined KPIs: Provide end-users with realistic measurements for their own KPIs for every security strategy and operational protocol for crowd management they wish to implement and evaluate that may be difficult or even impossible to obtain in real life without disrupting the everyday operational flow. Examples from the use of iCrowd for implementing risk-based security strategies at Border Crossing Points (BCPs) and evaluating their performance against rule-based security strategies in terms of security, operational efficiency and cost are provided in   .
The scenarios that were developed for the SAFETY4RAILS project involved the use of large complex 3D models with automatic escalators and turnstiles, the simulation of evacuations, and the use of cameras and guards to detect blind-spots and aid in the effectiveness of CCTV systems and patrol routes.
A very important feature that was used is the crowd and object detection by agents and other live entities such as cameras, which enables the user to design a CCTV system and assess its effectiveness as a crowd motion tracker in conjunction with an automated anomaly detection system    . iCrowd considers a realistic field-of-view for each camera based on its location and viewing angles, as well as the complexity of the geometry and even the real-time crowd congestion, which may affect the camera’s performance. Expanding on this feature, iCrowd offers detection evasion capabilities to its agents, enabling the user to simulate malicious actors who attempt to avoid being seen by cameras and mobile security personnel. In the figure below, one can see a screenshot of the iCrowd simulator showing a large number of agents inside a metro station, along with an area coverage heatmap.
iCrowd also offers information on pressure forces on agents due to high congestion levels. Pressure forces are generated by every agent with a non-zero velocity and are propagated in a cumulative way through the agents when an agent cannot react upon a received pressure (move away from it). iCrowd provides mechanisms for recording geospatial information throughout a simulation, such as congestion levels, pressure levels, and coverage by CCTV cameras and guards for each area of the simulated environment at each step of the simulation. Through a native OpenGL-based visualizer, this information can be inspected both during runtime and in post, in the form of heat-maps overlaid on top of the 3D environment. This can be seen in the figure below. The information can also be exported in common text files (CSV-formatted files) to enable post-processing by external software.
Figure 1 – Screenshot of multiple agents inside a station in iCrowd simulator
Figure 2 – Screenshot of a heatmap showing simulated congestion levels at a metro station
Interoperability with other SAFETY4RAILS tools
The iCrowd simulator features connectivity through HTTP, along with a dedicated implementation for connecting to an Apache Kafka server, such as SAFETY4RAILS’s Distributed Messaging System (DMS). DMS is the unified approach of tool intercommunication in SAFETY4RAILS, providing a software and platform agnostic communication layer that tools from heterogeneous environments can leverage. This enables iCrowd to integrate with other tools of the project, to provide a more informed and functionality-rich simulation platform, along with more useful metrics for the end-users to effectively evaluate their resilience strategies. Other tools can also receive information from iCrowd, such as movement trajectories, status updates about agents or assets, or results. Since DMS retains messages instead of just passing them through, it is possible to have both synchronous (real-time) and asynchronous integrations.
In the context of SAFETY4RAILS, the iCrowd simulator has been integrated with RINA’s own Bomb-Blast 3D (BB3D) tool. The goal was for iCrowd to realistically simulate the impact of a bomb explosion on the crowd and infrastructure, regarding injuries, fatalities, and disabled assets of the environment. BB3D provides an on-demand simulation platform that accepts a request for a bomb explosion, simulates it, and returns the results. The request includes the details of the bomb, its location in the predefined 3D geometry, and a summary of the crowd distribution around it. The results include the pressure levels due to the explosion and survival probability of a human at each point in space, as seen in the figure below.
Figure 3 – Screenshot showing the radius and impact severity of a simulated bomb blast through information retrieved from BB3D
As a result of this integration, a user of the iCrowd simulator can start a simulation scenario and dynamically define a place and time for a bomb to be detonated. They can even define multiple bombs to be detonated at various times. When the simulation reaches that point in time, iCrowd will automatically pause, triggering a request for the bomb explosion simulation to BB3D through DMS. When BB3D responds with the results of the simulation, they will be retrieved, applied, and the simulation will automatically continue. Based on the survival probabilities and infrastructure damages sent by BB3D, iCrowd will apply injury levels to nearby agents, and disable any relevant assets.
Further information regarding the use of the iCrowd simulator in SAFETY4RAILS can be found in the public Deliverable 5.2 where a detailed analysis of the functionality tested is presented. The iCrowd simulator was successfully demonstrated in the simulation exercises in Metro de Madrid and Ankara. The aspiration moving forward is to increase the adoption of simulation tools in the Railway sector in an attempt to improve even further the overall security and resilience.
-  https://cordis.europa.eu/project/id/241905
-  https://vimeo.com/51434409/648f3dbc54
-  https://vimeo.com/105113529/028e39b58f
-  V. Kountouriotis, S. C. Thomopoulos, and Y. Papelis, “An agent-based crowd behaviour model for real time crowd behaviour simulation,” Pattern Recognition Letters, vol. 44, pp. 30–38, 2014.
-  V. I. Kountouriotis, M. Paterakis, and S. C. Thomopoulos, “iCrowd: agent-based behavior modeling and crowd simulator,” in Signal Processing, Sensor/Information Fusion, and Target Recognition XXV, 2016, vol. 9842, p. 98420Q.
-  https://vimeo.com/223308809/4d0650bb33
-  https://vimeo.com/565219711/91a329fec0
-  https://vimeo.com/154047986/fb8b94cf62
-  https://vimeo.com/719864479/504362abb0
-  https://vimeo.com/716438202/da3b62f5ba/
-  https://vimeo.com/280690026/19dc5c86b9
-  https://vimeo.com/680373088/666854c5d5
-  https://vimeo.com/680311392/d4ee3557b6
-  https://vimeo.com/680319475/e2009fca58
-  Stelios C. A. Thomopoulos, Stelios Daveas and Antonios Danelakis, “Automated real-time risk assessment for airport passengers using a deep learning architecture,” Proceedings Volume 11018, Signal Processing, Sensor/Information Fusion, and Target Recognition XXVIII; 110180O (2019) https://doi.org/10.1117/12.2519857 Event: SPIE Defense + Commercial Sensing, 2019, Baltimore, Maryland, United States, 2019.
-  Giorgos Bouritsas, Stelios Daveas, Antonios Danelakis, Stelios C.A. Thomopoulos, “Automated Real-time Anomaly Detection in Human Trajectories using Sequence to Sequence Networks,” 2019 16th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS), p. 1-8, 2019/9/18.
-  Stelios C. A. Thomopoulos, “Risk Assessment and Automated Anomaly Detection Using a Deep Learning Architecture,” in Artificial Neural Networks and Deep Learning-Applications and Perspective, Editor IntechOpen, DOI: 10.5772/intechopen.96209, Published: February 18th 2021.
 Stelios C. A. Thomopoulos and Christos Kyriakopoulos “Anomaly detection with noisy and missing data using a deep learning architecture”, Proc. SPIE 11756, Signal Processing, Sensor/Information Fusion, and Target Recognition XXX, 117560R (16 April 2021); https://doi.org/10.1117/12.2589981 .
Authors: Stelios C. A. Thomopoulos, Christos Kyriakopoulos,Konstantinos Panou,
Integrated Systems Laboratory
Institute of Informatics & Telecommunications