Review of Virtual Simulation of Crowd Motion for Urban Emergency Management

Abstract

With increasing urbanization in China, there are more and more urban crowds, bringing new challenges to public safety management. The analysis and modeling of crowd behavior are important aspects of emergency management in smart cities. This paper reviews and summarizes existing research into crowd behavior simulation and analyzes existing simulation data collection methods. Crowd simulation technology is discussed from three aspects: crowd modeling, motion navigation, and emotion-driven crowd animation. The indicators of the simulation models are analyzed from the perspective of model evaluation. Finally, combined with the status quo, the existing research limitations are analyzed, and the direction of further research into crowd simulation behavior is discussed.

Keywords

crowd management emergency evacuation pedestrians simulation

The national urbanization process leads to more and more crowds in urban areas. Incidents involving public security are becoming more and more frequent. Failure to properly manage crowds in crowded places can lead to casualties and pose new challenges for public safety management. Crowd evacuation has become an important issue that cannot be ignored in emergency situations.

A group consists of two or more interacting individuals. A crowd formed by the mutual movement of multiple agents can be simulated on a computer. Virtual crowd animation can deduce various aspects of the crowd ( 1 , 2 ). Observing possible or existing crowd movements from different scales and perspectives will become an important means of crowd management and public education in the future. Crowd evacuation from buildings has always been a difficult feature in public safety management. If the crowd is divided into two roles—leader and follower—trained leaders play an important role in evacuation behavior. For example, different proportions of leaders in the crowd will lead to different evacuation results ( 3 ). Lu et al. ( 4 ) divided the crowd into civilians, terrorists, and heroes, and studied the self-organization of crowds. Serious games can induce human behavior. Combined with questionnaires and expert opinions, simulations of pedestrian evacuation from buildings can be constructed, illustrating human behavior when escaping from unknown buildings in an emergency ( 5 ).

China is vigorously promoting the construction of smart cities, making it urgent that the effectiveness of urban emergency management be improved. For example, Karamouzas et al.’s ( 6 ) research generates stable, collision-free, and smooth crowd movement under specific scenes and densities (Figure 1). Chen et al. ( 7 ) proposed a three-layer analysis framework for pedestrian evacuation simulations. Visual auxiliary analysis uses charts, virtual environments, crowd animations (2-D or 3-D), and other methods to help analyze the psychological changes in a crowd during an emergency. If this can be combined with physical models, it can better describe the scene in a crowd emergency evacuation, and deduce or predict the movement information of the crowd. Through visualization, people can quickly understand the trajectory and direction of a crowd’s movement, better observe the movement trend of the crowd, and greatly reduce the burden of people’s visual cognition. It is of great practical significance to study the crowd motion simulation for urban emergency management. Efficient and visual reproduction of crowd movement behavior in an emergency can be used to deduce and predict crowd behavior. This provides a means of visual auxiliary analysis for urban emergency safety management.

Figure 1.

Crowd movement scenario ( 6 ).

We searched “crowd simulation” in the Scopus literature database and collected 5,226 results. Topics related to “virtual” and “emergency” were further selected from these. Major journals that publish relevant representative literature include: Neurocomputing; ACM Transactions on Graphics; Safety Science; Computer Graphics Forum; Computers & Graphics; Autonomous Agents and Multi-Agent Systems; Physica A: Statistical Mechanics and its Applications; Visual Computer; IEEE Computer Graphics and Applications; IEEE Transactions on Visualization and Computer Graphics, Simulation Modeling Practice and Theory; IEEE Transactions on Systems, Man, and Cybernetics: Systems, Computer Animation, and Virtual Worlds; IEEE Transactions on Intelligent Transportation; Information Sciences; and other journals. In this paper, we select representative work from these journals and also draw on a small number of international conferences and other journals.

In this paper, we start from the needs of urban emergency management. We analyze the selected documents. This study targets the behavioral characteristics of crowds in the process of emergency evacuation, including the trajectory of movement, emergency evacuation time, evacuation speed, direction of movement, path selection, and so forth. At present, research into crowd behavior is mainly aimed at doing some coarse-grained analysis. The collection of motion data is the basis of crowd simulation research, and we focus on the behavioral characteristics and movement law of crowd evacuation under emergencies. Combined with the verification of the simulation model, the problems that may affect future virtual simulation of crowd behavior are put forward to provide reference for researchers in this field. The research framework of this paper is shown in Figure 2.

Figure 2.

Research framework.

Data-Driven Simulation

In the research of crowd movement, data collection is the key link. Basic data are not only the input data of motion simulation, but also the basis of establishing motion models. Their accuracy will directly affect the simulation results ( 8 , 9 ). Bernardini and Quagliarini ( 10 ) collected videos of recent terrorist acts all over Europe, which can be used to analyze and build a crowd evacuation model under terrorist acts. We analyze the acquisition of crowd motion simulation data from three aspects: video surveillance, evacuation drill, and unmanned aerial vehicle (UAV) filming.

Video Surveillance

At present, big data applications have been implemented in all walks of life. Video surveillance is a complete big data technology application process from front-end video technology to middle-stage mass storage to back-end big data analysis. Video surveillance is an important part of a security system. The traditional monitoring system includes front-end camera, transmission cable, and video monitoring platform, as shown in Figure 3.

Figure 3.

Monitoring system.

Video surveillance has long played an indisputably important role in various fields. Video surveillance is one of the important means of information acquisition. With the popularization of high-definition and intelligent technology, video surveillance has become a standard facility in major public places. There is constant innovation in the application technology of video surveillance systems. A large quantity of videos and images is generated, such as traffic big data, emergency crowd escape big data, and so forth, which provides scientific data support for crowd movement ( 11 ). To mine and simulate the real crowd behavior from video, Zhao et al. proposed a data-driven approach to simulate the realistic locomotion of virtual pedestrians. This extracts information from the video to simulate and avoid collisions. It also realizes hierarchical clustering with a distance function. The proposed model can generate realistic crowd behaviors ( 12 ). The distance function is a distance function of one pedestrian to another. The minimum mutual distance (MMD) between a neighbor $j$ and the subject pedestrian $i$ is $MMD (i, j)$ , during the time steps that $j$ is a neighbor of $i$ , as shown in Equation 1

\begin{matrix} M M D (i, j) = \min_{t^{'}} ‖ (x_{i}^{t^{'}} - y_{i}^{t^{'}}) - (x_{j}^{t^{'}}, y_{j}^{t^{'}}) ‖ \\ t' \in {t | j \in Neighbour (i, t), t > t_{c}} \end{matrix}

(1)

where $(x_{i}^{t'}, y_{i}^{t'})$ and $(x_{j}^{t'}, y_{j}^{t'})$ are the positions of $i$ and $j$ at time step $t'$ , respectively, and $Neighbour (i, t)$ is a set of pedestrians located in i’s vision area at $t$ .

The time step $t_{MMD} (i, j)$ when the subject pedestrian $i$ and neighbor $j$ reach their smallest mutual distance is shown in Equation 2:

\begin{matrix} t_{MMD} (i, j) = \arg min_{t'} ‖ (x_{i}^{t'}, y_{i}^{t'}) - ‖ (x_{j}^{t'} - y_{j}^{t'}) \\ t' \in {t | j \in Neighbour (i, t), t > t_{c}} \end{matrix}

(2)

The advantages of video surveillance are:

Video monitoring can monitor automatically with long data recording times. It helps to retrieve and use data repeatedly.

Monitoring data can be interconnected with many devices to realize data sharing.

The disadvantages of video surveillance are:

The data clarity needs to be improved. It is not recognizable by the human eye.

It needs sustained and stable power support, and has high energy requirements.

As a way to obtain data, existing video surveillance systems are more a means of evidence collection; they rarely give early warning of incidents.

Evacuation Drill

Catastrophic accidents are dangerous and are hard to replicate, making it difficult to obtain real accident data. In cases of fire, earthquake, building collapse, or other events, evacuating people away from the scene is the priority. The objective of evacuation is to minimize losses resulting from accidents and to ensure people’s safety. Researchers mostly obtain the empirical data of crowd evacuation processes through small-scale evacuation drills, using computers to simulate crowd behavior and evacuation processes. Organizing people to conduct evacuation drills in specific scenes is an effective way to obtain evacuation-related data in public gathering places such as subways and schools. Haghani and Sarvi ( 13 ) conducted an evacuation experiment, shown in Figure 4.

Figure 4.

Evacuation drill ( 13 ).

The main object of data collection of an evacuation drill is the time and space data of the evacuation, that is, to collect the time and route taken by each evacuee arriving at the designated location.

The main advantages of crowd evacuation drill are:

The scene is real and more in line with the actual situation of an emergency evacuation.

Various types of evacuation data can be obtained, such as evacuation route, evacuation time, and so forth.

Various exercises of setting scenarios (such as fire) and accident locations (such as a floor) can be carried out to obtain relevant data.

The main disadvantages of crowd evacuation drill are:

It is difficult to organize and implement an evacuation drill. It requires a lot of human, material, and financial resources. The security of people during the exercise is also a key consideration.

It is impossible to track and record the relevant attributes and evacuation trajectories of all the people at the same time.

The fidelity and execution of evacuation exercises cannot be compared with real emergencies.

Through repeated evacuation exercises, real crowd movement data can be obtained to guide people to use scientific methods to achieve effective escape in disasters.

UAV Shooting

AV shooting mainly uses a wireless monitoring network to monitor various large-scale scenes and continuously obtain data required by various applications, such as traffic data and data on crowd movements in an emergency ( 14 ). We can think of UAVs as flying sensors. They extend human vision and perception to the air and improve the acquisition of crowd movement data.

The advantages of UAV filming are:

It is affordable. The cost of aerial photography is greatly reduced with the use of UAVs.

It produces high data quality. At present, the data quality obtained by low altitude rotor aerial photography is much higher than the remote sensing data obtained by aerial photography. The 360° omni-directional filming can be carried out around the target.

However, UAV technology is not mature enough. The disadvantages of UAVs are:

The ability to avoid collisions needs to be improved.

The relay capacity of the battery is insufficient.

At present, the application of data collection and analysis based on UAVs are in their infancy. The field of crowd movement monitoring and data acquisition has not been fully extended to the field of consumers.

A comparison of various simulation data acquisition methods is shown in Table 1.

Table 1.

Comparison of Various Simulation Data Acquisition Methods

Data acquisition	Cost	Data quality	Application scenario	Application market
Video surveillance	Medium	Low	Small-scale	High penetration
Evacuation drill	High	High	Small-scale	High penetration
Unmanned aerial vehicle shooting	Low	High	Large-scale	Low penetration

The current market penetration rate of video surveillance and evacuation drills is high. Video surveillance and evacuation drills are only applicable to small-scale scenes. The economic cost of an evacuation drill is high, while the quality of the data obtained by video surveillance is low. UAV shooting is suitable for large-scale scenes and can obtain high data quality. From a long-term perspective, it has scope to develop extensively.

How to model the behavior of crowd movement with a data-driven method has become a key focus in crowd simulation. At present, there are two mainstream data-driven modeling methods. One is to minimize the defined rules. Jaklin et al. ( 15 ) measured the consistency and sociality of the crowd according to the existing empirical data of a real crowd and realized crowd simulation at the global and local levels.

Another data-driven method focuses more on the training model. It uses collected crowd behavior characteristics to train and generate the crowd behavior models of various agents, deploying excellent algorithms to calculate the parameters in the model. Zhong et al. ( 16 ) proposed a new data-driven modeling framework, the reciprocal velocity obstacle (RVO). They adopt the RVO2 to generate the microscopic collision avoidance behaviors in the bottom layer. The behaviors of new pedestrians appearing in each source region is modeled as a Poisson process (Equation 3). This models the goal selection patterns of pedestrians by using a probability matrix. The probability of pedestrians moving from the $i th$ source regions to the $j th$ destination regions is labeled as $p_{i, j}$ . The velocity field is used to control the moving directions of pedestrians. A 2-D vector $v_{i, j, k}$ is used to represent the future moving directions of the pedestrians in the grid $cell (j, k)$ . Taking the pedestrian trajectory of New York Grand Central Termina and ETH walking pedestrians as an example, the proposed framework is effective for generating crowd behaviors that are similar to those observed in the video for the crowd density distribution.

p {N_{i} (t + τ) - N_{i} (t) = k} = \frac{{(λ_{i} τ)}^{k} \exp (- λ_{i} τ)}{k!}

(3)

Bera et al. ( 17 ) obtained the crowd trajectory information from the video, including pedestrian position, current speed, and final target position, which is used to calculate the pedestrians in the crowd. The advantage is that it does not need many training examples for offline training and can directly use a k-means data clustering algorithm to group the trajectory behavior characteristics observed in a specific time window. These trajectory behavior features are used for motion segmentation, anomaly detection, and real data-driven simulation for virtual reality. A support vector machine can be used to train the navigation model and learn route selection ( 18 ). The machine can look at training samples from videos and when combined with a deep-neural-network can predict the characteristics of crowd modeling ( 19 – 22 ). Toledo et al. ( 23 ) combined data-driven and fuzzy logic to deal with the uncertainty of crowd movement.

Agent-Based Simulation

Crowd evacuation simulation can be applied to safety management and architectural design. It mainly focuses on how to model crowd movement in an emergency to avoid collisions in complex dynamic environments. From the published literature, many existing crowd motion models can be roughly divided into macro (based on continuum) and micro (based on agent) models as shown in Figure 5.

Figure 5.

Simulation model.

The macro model regards the crowd system as a whole, which is suitable for real-time simulation of a very large crowd. It usually follows the characteristics of flow. The main concern is that the overall crowd movement looks real. Micro models use complex cognitive models to study crowd movement from the perspective of individual behavior simulation. They mainly focus on individual autonomous behavior and its interactions. The micro model focuses on the individuals in the population. The basic principle is to explore the group movement rules by studying the behavior simulation of individuals. The micro model describes the group by effectively quantifying the individual behavior, and studies the relationship between the individual and the group from the perspective of the individual. There are discrete models and continuous models in micro models, and the most studied discrete model is the cellular automata model. Georgoudas et al. ( 24 ) proposed a group evacuation model based on cellular automata from the perspective of dynamics, as shown in Equation 4:

\vec{F} (\vec{r}) = \frac{q}{4 π ε_{0}} \sum_{i = 1}^{N} \frac{Q_{i} (\vec{r} - {\vec{r}}_{i})}{| \vec{r} - {\vec{r}}_{i} |} = \frac{q}{4 π ε_{0}} \sum_{i = 1}^{N} \frac{Q_{i}}{R_{i}^{2}} {\hat{R}}_{i} = q \hat{E} (\vec{r})

(4)

where $Q_{i}$ and ${\vec{r}}_{i}$ are the magnitude and position of the $i$ th charge, and $\hat{R}$ is the unit vector in the direction of ${\vec{R}}_{i} = \vec{r} - {\vec{r}}_{i}$ .

The continuous model can better describe the exit selection, the moving crowd, the crowded aisle, the congestion, and other phenomena of a crowd in an emergency evacuation. As a typical representative of the agent model ( 25 ), Helbing’s social force model (SFM) can simulate crowd congestion at the exit (Figure 6). The model mainly describes the interaction between individuals in the crowd. From the perspective of mechanics, an individual’s movement behavior is composed of the driving force to the target, the repulsive force to avoid others or objects, and the attraction of export (Equation 5).

m_{i} \frac{d v_{i}}{dt} = f_{i}^{driven} + \sum_{j (\neq i)} f_{ij} + \sum_{w} f_{iw}

(5)

where

$f_{i}^{driven}$ represents the driving force of the pedestrian by the target,

$f_{ij}$ represents the interaction force between the pedestrian $i$ and the pedestrian $j$ , and

$f_{iw}$ represents the interaction force between the pedestrian and the obstacle.

Figure 6.

Social force model of Helbing ( 25 ).

The application of macro and micro models in the field of crowd simulation has its own characteristics, which are summarized in Table 2.

Table 2.

Comparison Between Macro and Micro Models

Model	Data	Granularity	Features
Macroscopic	Whole	Rough	Simulate the crowd from the aspects of aggregation characteristics and flow characteristics
Microscopic	Individual	Fine	Study individual behavior simulation to explore the law of group movement

The emergence of agent-based simulation technology provides a new idea for the design and analysis of crowd evacuation schemes in emergencies. Observing and predicting crowd behavior in a three-dimensional and intuitive way is helpful to crowd anomaly detection and crowd behavior supervision and management ( 26 ). Micro analysis based on agents can simulate more detailed movement processes from the perspective of an individual. Ku et al. used a t-test to statistically analyze the means of transportation used by people and the number of people infected with the COVID-19 virus. The experimental results showed that the epidemic had a significant impact on the traffic mode of tourists. The use rate of private cars increased, and the public transport service rate showed a downward trend (such as bus and railway passenger transport) ( 27 ). D’Orazio ( 28 ) uses the agent method to model the transmission process of the virus, to analyze the effect of effective “social distance” on mitigating personal risks, and to evaluate the safety of tourist cities with large population mobility. The travel chain of the agent is shown in Table 3. Kim used a cellular automata model to study the movement of disabled people. Based on the 2-D grid, he analyzed the walking direction of pedestrians in forward, left and right directions (Figure 7). The average walking speed was significantly reduced. It is suggested that consideration be given to disabled pedestrians in the evacuation simulation ( 29 ). Ku et al. simulated how passengers encounter and infect each other on public transportation by tracking the movement of passengers. They used the smart card data to implement a traveler’s trip chain and performed traffic assignment under the agent-based model. The records of two databases, including the smart card data and real data on infected individuals, were adopted for constructing and analyzing the model. Information on smart card data includes the identifier ID of the transit user, location, boarding time, and disembarking time ( 30 ).

Table 3.

Travel Chain of Agent

Time	Activities
Morning/afternoon/evening	Randomly moving in the city areas by the movement-at-breaks value, to consider an equal probability of interacting with any other person within the world
Lunch/dinner	Moving near the initial generation position by trying to maintain the initial distance
Night sleep	Remaining at the initial generation position

Figure 7.

Pedestrian movement direction based on cellular automata ( 29 ).

Micro analysis based on agents can simulate more detailed movement processes from the perspective of the individual. Crowd evacuation simulation in emergencies mainly includes crowd modeling, motion navigation, and emotion-driven crowd animation, as shown in Figure 8.

Figure 8.

Crowd evacuation simulations.

Crowd Modeling

The autonomy, sociality, reflection and initiative of individuals in the agent model are suitable for simulating the interaction between individuals in a crowd. Agent technology can, therefore, be applied to crowd simulation. Trusted agents should be able to perceive the surrounding environment and adopt the most appropriate behavior in the current environment. How to construct individuals with realistic behavior is an important research direction in crowd modeling, and is becoming a focus of crowd research.

Mukai et al. proposed a virtual human behavior model. The model obtains information such as personal position and forward direction through vision and quantifies the big five personality models on seven levels. Each virtual human has two functions of search and collision avoidance ( 31 ). With the help of agent-based motion models (AMMs), the crowd motion types are identified from the crowd motion trajectory. The unknown noise in the crowd trajectory is filtered. The similarity with AMMs is measured to generate crowd motion characteristics. The synthetic data set “syncrowd” is given as the crowd training set ( 32 ). Martinez-Gil et al. analyzed pedestrian models, validation techniques, and multi-scale methods from the perspective of pedestrian dynamics. He classified existing pedestrian models and evaluated the behavioral quality of the simulation system. Facing multi-scale pedestrian modeling, he focused on the behavior model scale (the combination of micro and macro pedestrian models) and the second scale (from individuals to crowds) ( 33 ). For urban crowd simulations in large spaces, Sudkhot and Sombattheera proposed a multi-agent-based framework. They used belief desire intention (BDI) to model the behavior of individual agents, and RVO for agent navigation. They verified the feasibility of the framework through simulation steps, execution time, and visualization ( 34 ).

Movement Navigation

Motion navigation based on agents is an important part of crowd motion. Jin et al. ( 35 ) proposed a method based on vector fields to simulate crowd navigation through interactive control. Wang et al. proposed a framework for crowd formation through hierarchical planning, including cooperative tasks, coordinated behavior, and action control plans. The team achieved global path planning in cooperative tasks. Behavior planning was coordinated with a time-space table. A gaze model was combined with fuzzy logic control to realize agent action control planning ( 36 ). For persistent conflicts in computer games and crowd simulation, Banerjee proposed an improved conflict-based search, called symmetric ghost conflict, for multi-agent path finding ( 37 ). Han and Liu proposed a path selection model based on the evacuation path set. He created a sparse path set by using the social force model, then discretized and optimized the path. The method he proposed to disperse congestion does not consider the impact of pedestrian physiological and psychological differences on route selection during evacuation ( 38 ). Ren et al. ( 39 ) modeled individual psychological factors for path planning in emergencies (Figure 9). Combined with the surrounding environmental density, hazard intensity, and other information, he selected the path according to the individual’s cognitive domain information.

Figure 9.

Evacuation from a subway (emergency and path re-planning) ( 39 ).

In path planning, the shortest path is not always the best method for evacuation. The best method is related to the total number of simulated agents and their initial distribution in the virtual environment ( 40 ). Aiming at the motion trajectory and collision avoidance behavior of individuals at the corner, He et al. transformed the shadow obstacle model into optimal reciprocal collision avoidance (ORCA), half plane, and expected speed respectively. They manually generated shadow obstacles and effectively handled the situation of corners or being beside a non-closed wall. This method could be applied to the behavior simulation of large-scale people in complex scenes ( 41 ).

Most of the agents have their own motion directions in the navigation process. The agent mainly deals with the congestion by optimizing the route, avoiding collisions, changing the exit, and so on, thus causing the change in the movement route or speed.

The crowd behavior simulation is shown in Table 4. Many researchers use 3-D technology to visualize realistic crowd movements. For continuous models, RVO, SFM, and other technologies are often used for collision detection at the bottom layer. At the high level, the crowd navigation algorithm considers many factors, and the shortest path is often an important factor.

Table 4.

Crowd Behavior Simulation

Literature	Agent collision	Crowd navigation	Behavior selection	Tool
Mukai et al. ( 31 )	Personal space	Virtual human model with vision	Selﬁshness	C++
Sudkhot and Sombattheera ( 34 )	RVO	Shortest path	Belief Desire Intention	Unity3D
Han and Liu ( 38 )	RVO	Shortest path	Route choice model	C#
Ren et al. ( 39 )	RVO	Mental A*	Time-cost, safety of the path, and the crowd flow	Unity3D
Jin et al. ( 35 )	RBF	Vector ﬁelds	Anchor points	C++, MATLAB
Wang et al. ( 36 )	SFM	Shortest path	GMA and fuzzy logic control	Unity3D
He et al. ( 41 )	ORCA	Shortest path and security policy	Shadow obstacle and ORCA models	Unity3D

Note: RVO = reciprocal velocity obstacle; RBF = radial basis function; SFM = social force model; ORCA = optimal reciprocal collision avoidance.

For example, the following is the pseudo code of the shortest path algorithm in literature Sudkhot and Sombattheera ( 34 ).

Algorithm 1: Dijkstra’s Shortest Path Algorithm
Function Dijkstra(Graph, source): foreachvertex v in Graph do dist[vertex] : = infinity previous[vertex] : = undefined end dist[source] : = 0 Q : = the set of all nodes in Graph whileQ is notempty do u : = node in Q with smallest dist[] remove u from Q foreachneighbor v of u : do alt : = dist[u] +dist_between(u, v) ifalt < dist[v] then dist[v] : = alt previous[v] : = u else is not neighbor key end end end returnprevious[] End Function

Algorithm 1: Dijkstra’s Shortest Path Algorithm

Function Dijkstra(Graph, source): foreachvertex v in Graph do
dist[vertex] : = infinity
previous[vertex] : = undefined
end
dist[source] : = 0
Q : = the set of all nodes in Graph
whileQ is notempty do
u : = node in Q with smallest dist[]
remove u from Q
foreachneighbor v of u : do
alt : = dist[u] +dist_between(u, v)
ifalt < dist[v] then
dist[v] : = alt
previous[v] : = u
else
is not neighbor key
end
end
end
returnprevious[]
End Function

The main factors affecting crowd motion navigation are crowd density, agent interaction, and environmental factors.

(1) Crowd density: how to use crowd density information to guide the agent through the crowd is one of the key factors in navigation. Toll et al. ( 42 ) combine the distance information and crowd density information to form a weighted graph of the density values on the grid, as shown in Figure 10. This regularly re-plans the path according to the density to disperse the flow of people and avoid path congestion.

The continuous model of crowd movement and the simulation model of crowd dynamics can be combined to analyze the crowd from the perspective of physical force and interaction. It reasonably limits the density of the crowd by using the finite compressibility rule, which can realize large-scale high-density crowd simulations ( 43 , 44 ). Dickinson et al. studied the impact of agent density on user experience and behavior. They used virtual reality to create virtual experiences and explored crowd simulation and human behavior. The experimental results show that density significantly increases the negative effects ( 45 ). When the crowd density is very high, overcrowding can easily cause people to fall, which is also one of the most common dangerous phenomena in evacuations. According to the change of crowd density during evacuation, the pedestrian deceleration equation is conducive to formulating a safer evacuation plan ( 46 ). Density-based crowd animation has attracted the attention of more and more scholars. Density plays a very important role in individual path planning and behavior. The model could reproduce crowd behaviors such as lane formation. Density-based crowd animation will still be one of the important research directions in the future.

(2) Agent interaction: various interactions between agents, especially in emergencies, play an important role in formulating reasonable motion navigation. Wang et al. proposed an interactive multi-agent system based on an immersive virtual environment. Users can choose roles for control. Based on the interactive interface of gesture recognition, users can interact with virtual humans in real time ( 47 ). Bernardini et al. studied how people interact with other people in an environment changed by earthquake. The database established includes gradual evacuation behavior and amount of movement (speed, acceleration, and distance from obstacles). The pedestrian dynamic chart in earthquake emergencies shows that when the density values are equal, the speed and flow are higher ( 48 ).

The interaction between agents helps the agent quickly understand the evacuation information (such as event type, evacuation route, exit information, etc.). For the agents without prior knowledge of the environment in particular, agent interaction helps them to make a reasonable evacuation route ( 49 ).

(3) Environmental factors: the actual urban environment is of great significance for crowd movement, especially evacuation in emergencies ( 50 ). Zhou et al. ( 51 ) considered the impact of emergency signs on crowd evacuation in the environment. Environmental modeling and perception affect crowd evacuation. The crowd perceives the risk in the evacuation scene and shares the environmental information by updating the perception map ( 52 , 53 ). Zhou et al. ( 54 ) incorporated fuzzy logic into crowd evacuation simulations, as shown in Figure 11. They combine human experience and environmental information to simulate one-way and two-way flow of pedestrian.

Different roles of agents (such as managers familiar with the environment) will have different effects on the route selection in the evacuation process. Gou et al. ( 55 ) introduced a unified space to quantify multiple factors that may affect human movement. Considering the difference of group members’ familiarity with the environment, this could simulate the interaction behavior of individuals and groups.

Figure 10.

Path planning with density information (without density information and with density information) ( 42 ).

Figure 11.

Results of lane formation in bidirectional pedestrian flows through a hallway ( 54 ).

According to the residents’ understanding of the space, personal behavior, knowledge, and roles can be integrated into the simulated evacuation. Different personal behaviors and residents’ backgrounds and knowledge (such as space and previous emergency experience) will affect the evacuation results. Analyzing personal behavior and background knowledge helps to control the crowd and improve the evacuation effect ( 56 ). Therefore, the familiarity of evacuees with the environment and whether the environmental indication information is clear play an important role in the decision making in the navigation of crowd evacuation.

Actually, pedestrians are not completely isolated in the process of movement and have certain sociality. Usually, people tend to move with their relatives and friends, which is a common phenomenon in crowd movements. In the video surveillance, we can see that mutual aid escape is a common behavior in the face of terrorist attacks, especially among relatives, friends, and familiar people. It can actively help vulnerable groups evacuate together. Pedestrians keep relatively close contact with each other, which has a certain impact on gait, speed, and direction of movement. For example, as people tend to take care of the weakest, the speed of the slowest person becomes the coordinated evacuation speed. Because of this sociality, joint behavior creates a certain magnetism which makes pedestrians act together. We believe that magnetic functional parameters can be applied to pedestrians, which bind and attract sub-groupings of friends/family together in the model. This conforms to the psychological simulation of pedestrians based on sociology. Behavior is explicit, while psychology is implicit. Between psychology and behavior, they will influence each other, and implicit psychology dominates explicit behavior. Incorporating psychological factors into the evacuation model will make the pedestrian simulation process more realistic.

In addition, in the process of evacuation, individual distraction behavior will also affect perception and behavior, and then affect the overall behavior of the crowd.

Emotion-Driven Crowd Animation

The research into emotion-driven crowd animation involves computer vision, artificial intelligence, psychology, and other related theories and methods. Reynolds’ ( 57 ) bird flock animation is the earliest classic paper, which mainly simulates the aggregation of birds in the natural state. It realized path planning, collision avoidance, and real-time rendering. Small groups are an important concept in group animation. They are usually composed of family members or close associates. According to different topics, members leave and form groups automatically ( 58 ). Different group attributes, such as social groups or tourists and tour guides, can control the transformation of groups ( 59 ). In crowd animation, we often see the collective evacuation of small groups with close social relations.

In an emergency, the crowd was in a state of panic. Panic drives people to have unconventional behavior. The influence of emotion on behavior must be considered in crowd behavior simulation. It is, therefore, necessary to introduce the theory and method of effective computing into crowd motion simulation. Compared with the results of an individual emotion model, there are very few studies on the crowd emotion model. The key to the crowd emotion model is to solve the spread of emotion among individuals, that is, to describe the process of emotion infection. At present, the work of Bispo and Paiva (60) and Bosse et al. ( 61 ) can be seen and reported. They mainly focused on the absorption of emotions among individuals by using multi-agent tool NetLogo. According to the characteristics of dense crowd evacuation in emergencies, it is necessary to establish the emotional infection algorithm of the crowd. Considering an individual’s social attributes (such as experience, personality, gender, age, etc.), emotion-driven crowd movement is more consistent with the real crowd movement law.

How to maximize the role of positive emotions in crowd evacuation? Liu et al. ( 62 ) divided individual emotion into positive emotion, negative emotion, and unstable emotion. Emotional infection, collision avoidance, falling, and other behaviors were simulated to observe the process of crowd emotional infection from different perspectives. Experienced administrators play an important role in controlling crowd emotions. The number, timing, and location of administrators will affect the emotional changes of the whole crowd. Durupinar et al. ( 63 ) divided the crowd into audience and mob according to personality. The audience was a passive crowd, while the mob was an emotional and irrational active crowd. The system created could standardize different crowds. It parameterized the common attributes of the mob to create collective misconduct. Decision making based on emotion lead to many emergency behaviors. Users only needed to adjust personality parameters to achieve diversity (Figure 12) ( 64 ). Wu integrated emotions into crowd real-time path planning. He considered the differences of different personalities on path selection. Path planning was realized in combination with a global directed graph and local search, which was suitable for large-scale crowd evacuation ( 65 ).

Figure 12.

Two crowd scenarios representing expressive and acquisitive mobs ( 64 ).

Mao et al. analyzed the degree of intimacy between peers. They introduced the intimacy coefficient into emotional infection and made peer decisions in combination with task difficulty ( 66 ). Liu et al. ( 67 ) considered the impact of emotional recurrence on the process of emotional infection and constructed a cyber physical social emotional recurrence model, focusing on simulating the process of emotional infection. Xu et al. ( 68 ) proposed the concept of a danger field for panic emotion modeling in a multi-hazard environment. As an effective modeling method, a data-driven method could learn parameters from real videos to adjust, verify, and optimize group simulation and to analyze the propagation process of panic ( 69 , 70 ).

People in real emergencies are driven by subjective emotion. Emotion has a significant impact on evacuation. Emotion-driven animations are shown in Table 5. Most of the emotion-driven animations support panic emotions. Emotion will affect the navigation of agents. Emotion is mainly reflected in expression, speed increase, and so forth. Most emotion-driven behaviors contain the phenomenon of emotional infection. There are various crowd behaviors driven by emotion. Crowd behaviors in emergency situations include collision avoidance, path planning, evacuation, and so forth. For example, in route selection, individuals in panic will choose different exits around high-density crowds to escape from the dangerous area as soon as possible. Integrating emotional factors makes crowd animation more scientific and can better express crowd movement in an emergency. The existing crowd animation algorithms do not consider fully the emotional factors of the crowd and could not realistically show the diversified path selection. Emotion-driven crowd animation is an important development direction of crowd simulation.

Table 5.

Emotion-Driven Animation

Literature	Emotional model	Emotional type	Emotional expression
Durupınar et al. ( 64 )	OCC emotion model and PAD sentiment model	Happy, sad, angry, feaful	Facial expressions, postures
Mao et al. ( 66 )	OCC emotion model and PAD sentiment model	Calm, hope, fear, angry	Peer behaviors during emergency evacuation
Liu et al. ( 62 )	Perception-based	Panic	Speed increases and agent color
Wu et al. ( 65 )	Based on emotional path preference	Relaxed, calm, and impatient	Path planning and agent color
Liu et al. ( 67 )	CPS-REC model	Fear, happy, sad, angry, disgusted	Evacuation in the cyber-physical society
Başak et al. ( 69 )	Durupinar	Panic	Speed increases, escaping directions, and obstacle avoidance
Cornes et al. ( 70 )	The susceptibility and decay to fear emotions	Fear	Speed increases and agent color
Bosse et al. ( 61 )	ASCRIBE model	Fear	Speed increases
Durupinar et al. ( 63 )	OCEAN model	Panic	Speed increases
Xu et al. ( 68 )	Multihazard situations	Panic	Speed increases

Note: OCC = Ortony, Clorre, and Collins (64); PAD = Pleasure-Arousal-Dominance (64); CPS-REC = cyber-physical society oriented recurrent emotional contagion (67); ASCRIBE= agent-based social contagion regarding intentions, beliefs and emotions (61); OCEAN = openness, conscientiousness, extroversion, agreeableness, and neuroticism (63).

Algorithm 2: Update Compound Emotions
if appraisalFactor[Admiration] > 0 ∧ appraisalFactor[Joy] > 0 then appraisalFactor[Gratification] ← (appraisalFactor[Admiration] + appraisalFactor[Joy])/2; if appraisalFactor[Pride] > 0 ∧ appraisalFactor[Joy] > 0 then appraisalFactor[Gratitude] ← (appraisalFactor[Pride] + appraisalFactor[Joy])/2; if appraisalFactor[Shame] > 0 ∧ appraisalFactor[Distress] > 0 then appraisalFactor[Remorse] ← (appraisalFactor[Shame] + appraisalFactor[Distress])/2; if appraisalFactor[Reproach] > 0 ∧ appraisalFactor[Distress] > 0 then appraisalFactor[Anger] ← (appraisalFactor[Reproach] + appraisalFactor[Distress])/2;

Algorithm 2: Update Compound Emotions

if appraisalFactor[Admiration] > 0 ∧ appraisalFactor[Joy] > 0 then
appraisalFactor[Gratification] ← (appraisalFactor[Admiration] + appraisalFactor[Joy])/2;
if appraisalFactor[Pride] > 0 ∧ appraisalFactor[Joy] > 0 then
appraisalFactor[Gratitude] ← (appraisalFactor[Pride] + appraisalFactor[Joy])/2;
if appraisalFactor[Shame] > 0 ∧ appraisalFactor[Distress] > 0 then
appraisalFactor[Remorse] ← (appraisalFactor[Shame] + appraisalFactor[Distress])/2;
if appraisalFactor[Reproach] > 0 ∧ appraisalFactor[Distress] > 0 then
appraisalFactor[Anger] ← (appraisalFactor[Reproach] + appraisalFactor[Distress])/2;

For example, the following is the Pseudocode of an algorithm of update compound emotions in literature Durupinar et al. ( 64 ).

In addition, the crowd is not isolated. In the urban environment, it must involve interaction with motor vehicles, bicycles, and so forth. ( 71 ). How to effectively simulate the real process of vehicle–pedestrian interaction in mixed traffic is also an important link in crowd movement in urban emergency management (Figure 13) ( 72 ).

Figure 13.

Vehicle–pedestrian interaction scenario ( 72 ).

Verification of Simulation Model

Establishing a correct, reliable and effective simulation model is key to ensuring the high reliability of the simulation results. Therefore, the research into the theory and method of system simulation model establishment and verification has been much valued in the field of national simulation. The verification of a simulation model is mainly used to determine the correctness of the simulation system representing the real world. It focuses on the extent to which the simulation system reflects the real world. The verification focuses on the verification of the simulation results, which provides a basis to evaluate the effectiveness of the model system. The validation and evaluation of a crowd motion simulation model is usually based on real crowd data.

The multi-agent-based crowd simulation algorithm solves the combinatorial optimization of all parametric crowd simulation algorithms. It compares various indexes of reference data and simulation output. Simulation models include micro and macro indicators and various reference data (real data, such as trajectory, macro measurement, sketch, etc.) ( 73 ). Cassol et al. used crowd simulation to reproduce and evaluate the exit performance under specific scenarios. He proposed a crowd simulation tool CrowdSim, which was used to automatically reproduce the crowd behavior of building exits. Quantitative and qualitative methods for evaluating the software were given ( 74 ). Pax and Pavón proposed the agent architecture. This used the characteristics of the problem to achieve a balance between performance and flexibility. They realized perception, movement in the low-level behavior of the agent, and realized simulation decision making in high-level. The indoor environment crowd simulation was realized ( 75 ). Barut et al. evaluated whether the simulated 3-D human agent walked at a constant collision-free speed. He formed a walking illusion through virtual crowd video, combined with factors such as walking speed, crowd density, camera tilt angle, and camera distance ( 76 ).

Virtual reality is an effective tool. Evacuation experiments based on virtual reality can be used as alternative methods to study human behavior under extreme conditions, such as earthquake, fire, and so on. Lin constructed a virtual subway station for evacuation experiment. Under psychological pressure, the subjects chose to follow the flow direction of most people. At the same time, culture has a significant impact on the flow of people ( 77 ). Feng et al. built a virtual hospital to conduct evacuation experiments. Participants are often influenced by others, especially the views and behaviors of people with authority ( 78 ). Feng et al. simulated their experiment composed of four pathfinding tasks, automatically collected pedestrian walking trajectories, head movements and fixation points, and used various indicators to evaluate the effectiveness of the experiment ( 79 ).

The verification of the simulation model involves multiple indicators. From the existing evacuation model, competitive evacuation is the theoretical basis of the simulation model. The commonly used verification indicators include evacuation time, degree of exit congestion, and evacuation trajectory.

(1) Evacuation time: evacuation time refers to the time it takes for people to escape in an orderly way from the dangerous area to the safe area in case of an emergency (such as fire, earthquake, terrorist attack, etc.). At present, evacuation time has become an important index to verify the effectiveness of most evacuation models. Evacuation time is affected by many factors, such as a complete evacuation process, familiarity with evacuation environment, whether people have received relevant rehearsal training, and so forth. The maximum evacuation time and average evacuation time are usually used to comprehensively verify the effectiveness of the model.

(2) Degree of exit congestion: urban public places are often crowded with a lot of movement. When an emergency occurs, it is difficult to evacuate safely in a short time. It is easy to cause crowd congestion, even stampedes, at an exit, resulting in many casualties. A reasonable exit selection model (especially for scenes with obstacles and uneven crowd distribution) can effectively avoid crossing the crowd and overcrowding at the exit. The congestion degree of an emergency exit is, therefore, also an important indicator of safe evacuation.

(3) Evacuation trajectory: in emergency situations, the reasonable evacuation trajectory of the crowd can reflect the effective evacuation process. In the process of emergency evacuation, affected by the surrounding environment, evacuees will be in a state of high panic, which will affect their ability to find a suitable path. It is difficult to calmly analyze and follow the evacuation instructions. Therefore, the route complexity of an evacuation path should also be used as an important index to verify the emergency evacuation model. Two kinds of judgments are usually made with an evacuation trajectory: one is to judge whether the simulated trajectory is consistent with the actual evacuation route selection; second, judge whether the straightness and curvature of the trajectory are consistent with the actual evacuation route.

There is no unified statement on the verifiable part of the simulation results. Most models use the evacuation time as the verification index, and some models use the trajectory or the degree of exit congestion to verify. In fact, the verification of a crowd simulation model in an emergency should integrate the above indicators, rather than using a single verification index. We hope to have as many parameters as possible to comprehensively verify the simulation results, such as motion trajectory, evacuation time, and exit congestion degree. We should pay attention to the process simulation of the crowd, not only to the time points at both ends. The simulation process is also consistent as far as possible with the verification parameters, both on the time axis–related nodes and on the trajectory. Integrating various verification indexes can more comprehensively and scientifically evaluate the feasibility of the model and further improve the accuracy of simulation.

Problems and Challenges

Crowd motion virtual simulation technology involves physics, cognitive science, human behavior, computing, and other related disciplines. After decades of rapid development, some research results have been accumulated in crowd evacuation simulation and crowd psychology modeling. Some corresponding crowd simulation software has been developed. For example, EVACNET, developed by University of Florida, is numerical simulation software for personnel evacuation. It focuses on simulating the evacuation time required by people. BuildingExodus is building-oriented personnel evacuation software developed by the University of Greenwich, which can be used to simulate evacuation behavior and different exit flows. However, a review of the literature suggests that the existing simulation results could not fully reflect crowd motion in emergencies, so there are still many problems and challenges in crowd motion simulation.

(1) Physical reality simulation of evacuation environment: the challenge is analyzed by observing the real video monitoring. From the real video surveillance, we find that there is a big gap between the existing simulation scenario and the actual situation. The existing evacuation simulation scenarios are relatively simple and often lack the description of the physical effects of the evacuation environment, which makes the process description of evacuation simulation less realistic. Zhuang et al. ( 80 ) describe the physical effects in earthquakes, which can better present evacuation scenarios. Research into crowd evacuation simulation needs to consider the physical effects of people and environment, which is a challenging direction of study.

(2) Data acquisition means: the existing data acquisition means are not perfect. Most of the safety evacuation exercises are arranged at known times and places, which are not sudden or random. The participants lack the ability to respond to emergencies. The authenticity of the collected crowd movement data needs to be improved. Existing video surveillance can only observe the crowd in the current limited area. The visual field of video surveillance is fixed and inflexible. It could not make “visual prediction” of the crowd’s behavior in advance. UAV can flexibly observe from multiple angles and quickly transmit data, but the air line of UAV needs to be improved. Therefore, as a data acquisition method, UAV shooting needs to be popularized.

(3) Model parameters: crowd behavior simulation is the most complex and difficult aspect of simulating evacuation processes. Not all evacuation behavior characteristics can be fully recognized or fully quantified. So far, there is no perfect model that can completely solve all aspects of human evacuation behavior. The simple model with few parameters has a single function. For example, the model considering only evacuation time is not comprehensive enough. The model with many parameters has relatively good functions, such as the model considering evacuation time and exit congestion. How to balance the performance and model parameters of the evacuation model is still the goal of model optimization.

(4) Computational complexity: the computational complexity of the existing crowd motion model is high. The current crowd motion navigation focuses on the direction, speed, and collision avoidance of individual motion. When the scale of a scene and its population is large, the computational complexity is still high. Although ordinary workstations already have GPU processing capacity, from the perspective of visualization, the complexity of drawing and rendering highly realistic crowd animation is still high. How to improve the efficiency of crowd animation algorithms so that they can simulate large-scale realistic crowd motion remains to be studied further.

(5) Verification method: crowd motion simulation lacks verification of the model. At present, most verification methods are compared with video data. The simulation experiment is not completely consistent with crowd evacuation behavior in emergencies. For example, there are different degrees of difference in the time to reach the exit and crowd evacuation routes. It is difficult to compare crowd simulations with real video. At present, the crowd movement from the video can be sumulated to some extent. According to the time deduction, the overall situation is similar. By optimizing the algorithm, the motion trajectory of the simulated crowd is as close as possible to the motion trajectory of the crowd in the video. Compare the monitored crowd at the intersection with the simulated crowd. The optimization technology is used mainly to adjust the simulation parameters. The goal is to minimize the gap between the moving position of the agent and the position of the real person in each time step, and to adjust the simulation parameters to the most appropriate state through continuous fitting training.

The statistical analysis of crowd simulation is based on multiple simulations, taking the average of experimental data. There will be appropriate fluctuations in each simulation, such as the initial position of the crowd, resulting in different simulation results such as evacuation time. Therefore, from the perspective of rationality and comprehensiveness, statistical verification is mainly applied.

More and more researchers have begun to pay attention to data-driven methods. With the help of advanced intelligent equipment, the collection, analysis, modeling, experiment, and verification of motion data have become a series of indispensable links, especially verification. At present, we have only adopted some indicators for comprehensive verification. How to achieve rapid and effective verification will be one of the future challenges.

Researchers have been able to obtain the motion information of the crowd such as the trajectory through video surveillance and other methods. However, the fine-grained gait analysis, expression, posture and other details have not been studied in depth, and there is significant room for improvement in the future. For example, this can describe the details of individuals in a fine-grained manner, and also include various actions. This can also transfer crowd behaviors through machine learning, such as learning the gait and speed of pedestrians.

Conclusions

The technology for simulating crowd behavior in emergencies has been the focus of much work in recent years in the field of crowd research. The simulation goal is to achieve efficient and realistic crowd movement, which can be applied to public safety management, military exercises, traffic design, architectural design, and other fields. Focusing on the needs of smart city emergency management, combined with effective simulation data collection methods (video surveillance, evacuation exercise, UAV shooting), this paper classifies and introduces the model and simulation animation of crowd behavior. We focus on existing agent-based crowd simulation technology. The overall research framework and ideas show that emotion-driven virtual simulation technology is an effective method for studying crowd movement. Combined with the verification of the simulation model, aiming at the shortcomings of the existing research, we suggest some problems needing further study. The solutions to these problems will contribute to the continuous improvement of crowd motion simulation technology, so as to make the crowd animation more realistic in emergencies.

Despite the rapid development of virtual simulation technology, it is still difficult to simulate group behavior realistically. The existing crowd simulation research is far from enough. Whether it is fine-grained gait, expression analysis, or verification methods (especially the comparison with real video), it has not yet achieved in-depth research and only stays on the surface. There is great room for improvement in the future to achieve a more realistic virtual crowd. The current crowd behavior model lacks a solid psychological theoretical basis, and there is still much room to improve its credibility. Emotion plays an indispensable role in group behavior. In particular, research into crowd behavior in emergency situations urgently needs to incorporate emotional factors. Combining psychological models to model crowd behavior in emergency situations may prove to be a fruitful area for future research.

Footnotes

Author Contributions

The authors confirm contribution to the paper as follows: study conception and design: C. Liu; data collection: Y. Chai; analysis and interpretation of results: Z. Liu; draft manuscript preparation: Y. Chai. All authors reviewed the results and approved the final version of the manuscript.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Ningbo Science Technology Plan projects (Grant No. 2022Z077), and Project of Zhejiang Provincial Department of Education (Grant No. Y202045394).

References

Thalmann

Musse

S. R.

Crowd Simulation, 2nd ed. Berlin/Heidelberg, German: Springer-Verlag, 2012.

M. L.

Jiang

Jin

X. G.

Deng

Crowd Simulation and its Applications: Recent Advances. Journal of Computer Science and Technology, Vol. 29, No. 5, 2014, pp. 799–811.

Pelechano

Badler

N. I.

Modeling Crowd and Trained Leader Behavior During Building Evacuation. IEEE Computer Graphics & Applications, Vol. 26, No. 6, 2006, pp. 80–86.

Yang

Zhang

Swarm Intelligence, Social Force and Multi-Agent Modeling of Heroic Altruism Behaviors Under Collective Risks. Knowledge-Based Systems, Vol. 214, 2021, p. 106725.

Almeida

J. E.

Rossetti

R. J. F.

Jacob

J. T. P. N.

Faria

B. M.

Coelho

A. L.

Serious Games for the Human Behaviour Analysis in Emergency Evacuation Scenarios. Cluster Computing, Vol. 20, No. 1, 2017, pp. 707–720.

Karamouzas

Sohre

Narain

Guy

S. J.

Implicit Crowds: Optimization Integrator for Robust Crowd Simulation. ACM Transactions on Graphics, Vol. 36, No. 4, 2017, Article 136.

Chen

Shi

Pedestrian Evacuation Simulation in Indoor Emergency Situations: Approaches, Models and Tools. Safety Science, Vol. 142, 2021, p. 105378.

Qiao

Yoon

Kapadia

Pavlovic

The Role of Data-Driven Priors in Multi-Agent Crowd Trajectory Estimation. Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 32, No. 1. https://doi.org/10.1609/aaai.v32i1.11593.

Bisagno

Conci

Zhang

Data-Driven Crowd Simulation. Proc., 14th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS), Lecce, Italy, IEEE, New York, 2017, pp. 1–6. https://doi.org/10.1109/AVSS.2017.8078494.

10.

Bernardini

Quagliarini

Terrorist Acts and Pedestrians’ Behaviours: First Insights on European Contexts for Evacuation Modelling. Safety Science, Vol. 143, 2021, p. 105405.

11.

Lee

K. H.

Choi

M. G.

Hong

Lee

Group Behavior From Video: A Data-Driven Approach to Crowd Simulation. Proc., ACM SIGGRAPH/Eurographics Symposium on Computer Animation, ACM Press, New York, NY, 2007, pp. 109–118.

12.

Zhao

Cai

Turner

S. J.

CLUST: Simulating Realistic Crowd Behaviour by Mining Pattern From Crowd Videos. Computer Graphics Forum, Vol. 37, No. 1, 2018, pp. 184–201.

13.

Haghani

Sarvi

Simulating Dynamics of Adaptive Exit-Choice Changing in Crowd Evacuations: Model Implementation and Behavioural Interpretations. Transportation Research Part C: Emerging Technologies, Vol. 103, 2019, pp. 56–82.

14.

Wang

A. W.

Fan

H. N.

Fang

D. Y.

UAV Data Collection Method Based on Data Value. Journal of Software, Vol. 28, No. 12, 2017, pp. 3206–3222.

15.

Jaklin

Kremyzas

Geraerts

Adding Sociality to Virtual Pedestrian Groups. Proc., 21st ACM Symposium on Virtual Reality Software and Technology, VRST ’15, ACM Press, New York, NY, 2015, pp. 163–172.

16.

Zhong

Cai

Luo

Zhao

Learning Behavior Patterns From Video for Agent-Based Crowd Modeling and Simulation. Autonomous Agents and Multi-Agent Systems, Vol. 30, No. 5, 2016, pp. 990–1019.

17.

Bera

Kim

Manocha

Interactive Crowd-Behavior Learning for Surveillance and Training. IEEE Computer Graphics and Applications, Vol. 36, No. 6, 2016, pp. 37–45.

18.

Tan

S. K.

Cai

A Data-Driven Path Planning Model for Crowd Capacity Analysis. Journal of Computational Science, Vol. 34, 2019, pp. 66–79.

19.

Song

Han

Sun

Zhang

A Data-Driven Neural Network Approach to Simulate Pedestrian Movement. Physica A: Statistical Mechanics and its Applications, Vol. 509, 2018, pp. 827–844.

20.

Haworth

M. B.

Berseth

Pavlovic

Faloutsos

Kapadia

Heterogeneous Crowd Simulation Using Parametric Reinforcement Learning. IEEE Transactions on Visualization and Computer Graphics, 2021. https://doi.org/10.1109/TVCG.2021.3139031.

21.

Oshita

Agent Navigation Using Deep Learning With Agent Space Heat Map for Crowd Simulation. Computer Animation and Virtual Worlds, Vol. 30, No. 3–4, 2019, p. e1878.

22.

Wei

Zhu

Xing

Learning Motion Rules From Real Data: Neural Network for Crowd Simulation. Neurocomputing, Vol. 310, 2018, pp. 125–134.

23.

Toledo

Rivalcoba

Rudomin

Fuzzy and Data-Driven Urban Crowds. Computational Science – ICCS, Vol. 10862, 2018, pp. 280–290.

24.

Georgoudas

I. G.

Kyriakos

Sirakoulis

G. C.

Andreadis

I. T.

An FPGA Implemented Cellular Automaton Crowd Evacuation Model Inspired by the Electrostatic-Induced Potential Fields. Microprocessors & Microsystems, Vol. 34, No. 7/8, 2010, pp. 285–300.

25.

Helbing

Farkas

Vicsek

Simulating Dynamical Features of Escape Panic. Nature, Vol. 407, No. 6803, 2000, pp. 487–490.

26.

Charalambous

Karamouzas

Guy

S. J.

Chrysanthou

A Data-Driven Framework for Visual Crowd Analysis. Computer Graphics Forum, Vol. 33, No. 7, 2014, pp. 41–50.

27.

D. G.

J. S.

Byon

Y. J.

Kim

J. Y.

Lee

S. J.

Changes in Passengers’ Travel Behavior due to COVID-19. Sustainability, Vol. 13, 2021, p. 7974.

28.

D’Orazio

Bernardini

Quagliarini

Sustainable and Resilient Strategies for Touristic Cities Against COVID-19: An Agent-Based Approach. Safety Science, Vol. 142, 2021, p. 105399.

29.

Kim

Ahn

Lee

Modeling Handicapped Pedestrians Considering Physical Characteristics Using Cellular Automaton. Physica A: Statistical Mechanics and its Applications, Vol. 510, 2018, pp. 507–517.

30.

Yeon

Lee

Hwang

Y. C.

Wong

S. C.

Safe Traveling in Public Transport Amid COVID-19. Science Advances, Vol. 7, 2021, p. eabg3691.

31.

Mukai

Tanaka

Chang

Crowd Simulation by Applying Individual Human Model With Vision. Proc., International Conference on Cyberworlds, Visby, Sweden, IEEE Computer Society Press, Washington, D.C., 2015, pp. 210–215. https://doi.org/10.1109/CW.2015.14.

32.

Liu

Lau

R. W. H.

Wang

Manocha

Exemplar-AMMs: Recognizing Crowd Movements From Pedestrian Trajectories. IEEE Transactions on Multimedia, Vol. 18, No. 12, 2016, pp. 2398–2406.

33.

Martinez-Gil

Lozano

García-Fernández

Fernandez

Modeling, Evaluation, and Scale on Artificial Pedestrians. ACM Computing Surveys, Vol. 50, No. 5, 2017, pp. 1–35.

34.

Sudkhot

Sombattheera

A Crowd Simulation in Large Space Urban. Proc., International Conference on Information Technology, Khon Kaen, Thailand, IEEE Computer Society Press, Washington, D.C., 2018, pp. 1–8. https://doi.org/10.23919/INCIT.2018.8584878.

35.

Jin

Wang

C. C. L.

Huang

Zhang

Interactive Control of Real-Time Crowd Navigation in Virtual Environment. IEEE Computer Graphics and Applications, Vol. 28, No. 6, 2008, pp. 37–46.

36.

Wang

Liu

Zhou

Cheng

Miao

Thalmann

N. M.

Crowd Formation via Hierarchical Planning. Proc., 15th ACM SIGGRAPH Conference on Virtual-Reality Continuum and its Applications in Industry, ACM Press, New York, NY, 2016, pp. 251–260.

37.

Banerjee

Davis

C. E.

Multiagent Path Finding With Persistence Conflicts. IEEE Transactions on Computational Intelligence and AI in Games, Vol. 9, No. 4, 2017, pp. 402–409.

38.

Han

Y. B.

Liu

Research on Route Choice Model Based on Evacuation Route Set and its Application in Crowd Evacuation Simulation. Chinese Journal of Computers, Vol. 41, No. 12, 2018, pp. 2653–2669.

39.

Ren

Gai

Peng

Ego-Centered Path Planning in Evacuation Simulation. Journal of Computer-Aided Design & Computer Graphics, Vol. 27, No. 9, 2015, pp. 1775–1785.

40.

Cassol

V. J.

Testa

E. S.

Jung

C. R.

Usman

Faloutsos

Berseth

Kapadia

Badler

N. I.

Musse

S. R.

Evaluating and Optimizing Evacuation Plans for Crowd Egress. IEEE Computer Graphics and Applications, Vol. 37, No. 4, 2017, pp. 60–71.

41.

G. Q.

Jiang

D. X.

Jin

Chen

X. J.

M. L.

Crowd Behavior Simulation Based on Shadow Obstacle and ORCA Models. Scientia Sinica Informationis, Vol. 48, No. 3, 2018, pp. 233–247.

42.

Toll

W. G. V.

A. F. C.

Geraerts

Real-Time Density-Based Crowd Simulation. Computer Animation and Virtual Worlds, Vol. 23, No.1, 2012, pp. 59–69.

43.

Sun

L. B.

Sun

X. F.

Qin

W. H.

Dense Crowd Simulation Based on Continuum Model and Aggregate Dynamics Model. Chinese Journal of Computers, Vol. 39, No. 7, 2016, pp. 1375–1392.

44.

Kim

Guy

S. J.

Hillesland

Zafar

Gutub

A. A. A.

Manocha

Velocity-Based Modeling of Physical Interactions in Dense Crowds. Visual Computer, Vol. 31, No. 5, 2015, pp. 541–555.

45.

Dickinson

Gerling

Hicks

Murray

Shearer

Greenwood

Virtual Reality Crowd Simulation: Effects of Agent Density on User Experience and Behavior. Virtual Reality, Vol. 23, No. 1, 2019, pp. 19–32.

46.

Kim

Han

Analysis of Evacuation Simulation Considering Crowd Density and the Effect of a Fallen Person. Journal of Ambient Intelligence and Humanized Computing, Vol. 10, No. 12, 2019, pp. 4869–4879https://doi.org/10.1007/s12652-019-01184-7.

47.

Wang

Dubey

Magnenat-Thalmann

Thalmann

An Immersive Multi-Agent System for Interactive Applications. Visual Computer, Vol. 29, No. 5, 2013, pp. 323–332.

48.

Bernardini

Quagliarini

D’Orazio

Towards Creating a Combined Database for Earthquake Pedestrians’ Evacuation Models. Safety Science, Vol. 82, 2016, pp. 77–94.

49.

Kullu

Güdükbay

Manocha

ACMICS: An Agent Communication Model for Interacting Crowd Simulation. Autonomous Agents and Multi-Agent Systems, Vol. 31, No. 6, 2017, pp. 1403–1423.

50.

Liu

Zhang

Modeling the Dynamics of Pedestrian Evacuation in a Complex Environment. Physica A: Statistical Mechanics and its Applications, Vol. 585, 2022, p. 126426.

51.

Zhou

Dong

Wang

Modeling and Simulation of Crowd Evacuation With Signs at Subway Platform: A Case Study of Beijing Subway Stations. IEEE Transactions on Intelligent Transportation Systems, Vol. 23, No. 2, 2022, pp. 1492–1504.

52.

Lovreglio

Ronchi

Nilsson

An Evacuation Decision Model Based on Perceived Risk, Social Influence and Behavioural Uncertainty. Simulation Modelling Practice and Theory, Vol. 66, 2016, pp. 226–242.

53.

Zafar

Zia

Muhammad

Ferscha

An Agent-Based Model of Crowd Evacuation Integrating Agent Perception and Proximity Pressure. Proc., 14th International Conference on Advances in Mobile Computing and Multi Media, Singapore, ACM Press, New York, NY, 2016, pp. 12–19.

54.

Zhou

Dong

Wang

F. Y.

Wang

Yang

Modeling and Simulation of Pedestrian Dynamical Behavior Based on a Fuzzy Logic Approach. Information Sciences, Vol. 360, 2016, pp. 112–130.

55.

Gou

Zhang

Simulation of Small Social Group Behaviors in Emergency Evacuation Based on Information Asymmetry. Journal of Computer-Aided Design & Computer Graphics, Vol. 30, No. 3, 2018, pp. 524–530.

56.

Chu

M. L.

Law

K. H.

Incorporating Individual Behavior, Knowledge, and Roles in Simulating Evacuation. Fire Technology, Vol. 55, 2019, pp. 437–464.

57.

Reynolds

Flocks, Herds and Schools: A Distributed Behavioral Model. ACM SIGGRAPH Computer Graphics, Vol. 21, No. 4, 1987, pp. 25–34.

58.

Karimaghalou

Bernardet

Dipaola

A Model for Social Spatial Behavior in Virtual Characters. Computer Animation and Virtual Worlds, Vol. 25, No. 3–4, 2014, pp. 505–517.

59.

Ren

Charalambous

Bruneau

Peng

Pettré

Group Modeling: A Unified Velocity-Based Approach. Computer Graphics Forum, Vol. 36, No. 8, 2017, pp. 45–56.

60.

Bispo

Paiva

A Model for Emotional Contagion Based on the Emotional Contagion Scale. Proc., 3rd International Conference on Affective Computing and Intelligent Interaction and Workshops, Amsterdam, Netherlands, IEEE Computer Society Press, Washington, D.C., 2009, pp. 1–6.

61.

Bosse

Hoogendoorn

Klein

M. C. A.

Treur

van der Wal

C. N.

van Wissen

Modelling Collective Decision Making in Groups and Crowds: Integrating Social Contagion and Interacting Emotions, Beliefs and Intentions. Autonomous Agents and Multi-Agent Systems, Vol. 27, No. 1, 2013, pp. 52–84.

62.

Liu

Hsu

H. H.

Chai

A Perception-Based Emotion Contagion Model in Crowd Emergent Evacuation Simulation. Computer Animation and Virtual Worlds, Vol. 29, No. 3–4, 2018, p. e1817.

63.

Durupinar

Pelechano

Allbeck

J. M.

Gudukbay

Badler

N. I.

How the Ocean Personality Model Affects the Perception of Crowds. IEEE Computer Graphics and Applications, Vol. 31, No. 3, 2009, pp. 22–31.

64.

Durupinar

Gudukbay

Aman

Badler

N. I.

Psychological Parameters for Crowd Simulation: From Audiences to Mobs. IEEE Transactions on Visualization and Computer Graphics, Vol. 22, No. 9, 2016, pp. 2145–2159.

65.

Real-Time Emotion-Integrated Crowd Path Planning. Journal of Image and Graphics, Vol. 20, No. 9, 2015, pp. 1262–1274.

66.

Mao

Fan

Zhao

Zhang

An Emotional Contagion Based Simulation for Emergency Evacuation Peer Behavior Decision. Simulation Modelling Practice and Theory, Vol. 96, 2019, p. 101936.

67.

Liu

Zhang

Hong

Liu

Recurrent Emotional Contagion for the Crowd Evacuation of a Cyber-Physical Society. Information Sciences, Vol. 575, 2021, pp. 155–172.

68.

Xie

Niu

Wang

Zhu

Deng

Zhou

Crowd Behavior Simulation With Emotional Contagion in Unexpected Multihazard Situations. IEEE Transactions on Systems, Man, and Cybernetics: Systems, Vol. 51, No. 3, 2021, pp. 1567–1581.

69.

Başak

A. E.

Güdükbay

Durupınar

Using Real Life Incidents for Creating Realistic Virtual Crowds With Data-Driven Emotion Contagion. Computers & Graphics, Vol. 72, 2018, pp. 70–81.

70.

Cornes

F. E.

Frank

G. A.

Dorso

C. O.

Fear Propagation and the Evacuation Dynamics. Simulation Modelling Practice and Theory, Vol. 95, 2019, pp. 112–133.

71.

Chao

Deng

Xiao

Miao

Jin

Dictionary-Based Fidelity Measure for Virtual Traffic. IEEE Transactions on Visualization and Computer Graphics, Vol. 26, No. 3, 2020, pp. 1490–1501.

72.

Han

Chao

Jin

A Simplified Force Model for Mixed Traffic Simulation. Computer Animation and Virtual Worlds, Vol. 32, 2021, p. e1974.

73.

Wolinski

Guy

S. J.

Olivier

A.-H.

Lin

Manocha

Pettré

Parameter Estimation and Comparative Evaluation of Crowd Simulations. Computer Graphics Forum, Vol. 33, No. 2, 2014, pp. 303–312.

74.

Cassol

Oliveira

Musse

S. R.

Badler

Analyzing Egress Accuracy Through the Study of Virtual and Real Crowds. Proc., IEEE Virtual Humans and Crowds for Immersive Environments (VHCIE), Greenville, SC, IEEE Computer Society Press, Washington, D.C., 2016. https://doi.org/10.1109/VHCIE.2016.7563565.

75.

Pax

Pavón

Agent Architecture for Crowd Simulation in Indoor Environments. Journal of Ambient Intelligence and Humanized Computing, Vol. 8, No. 2, 2017, pp. 205–212.

76.

Barut

Haciomeroglu

Sezer

E. A.

Perceptual Evaluation of Maneuvering Motion Illusion for Virtual Pedestrians. The Visual Computer, Vol. 4, 2018, pp. 1119–1128.

77.

Lin

Zhu

Becerik-Gerber

Do People Follow the Crowd in Building Emergency Evacuation? A Cross-Cultural Immersive Virtual Reality-Based Study. Advanced Engineering Informatics, Vol. 43, 2020, p. 101040.

78.

Feng

González

V. A.

Trotter

Spearpoint

Thomas

Ellis

Lovreglio

How People Make Decisions During Earthquakes and Post-Earthquake Evacuation: Using Verbal Protocol Analysis in Immersive Virtual Reality. Safety Science, Vol. 129, 2020, p. 104837.

79.

Feng

Duives

D. C.

Hoogendoorn

S. P.

Development and Evaluation of a VR Research Tool to Study Wayfinding Behavior in a Multi-Story Building. Safety Science, Vol. 147, 2022, p. 105573.

80.

Zhuang

Liu

T. T.

Huang

C. C.

Chai

Y. J.

Implicit Sentiment Analysis Based on Multi-Feature Neural Network Model. Soft Computing: A Fusion of Foundations, Methodologies and Applications, Vol. 26, 2022, pp. 635–644.