interactive behavior modeling for large-scale crowd...

Interactive Behavior Modeling for Large-Scale Crowd Simulations

Ming C. Lin

Dinesh ManochaDepartment of Computer Science

University of North CarolinaChapel Hill, NC 27599-3175919-962-1974, 919-962-1749

[email protected], [email protected]

Latika EifertAngel Rodriguez

Army Research Laboratory HRED STTCSFC Paul Ray Smith, Simulation and Training Technology Center

12423 Research Parkway, Orlando, FL 32826407-384-5338, 407-208-3009

[email protected],[email protected]

Keywords:multi-agent simualtion, emergent behaviors, collision-avoidance, least-effort crowds

ABSTRACT: We survey some of our recent works on real-time simulation of large-scale crowd movement. Our primaryfocuses include techniques for automatically computing collision-free trajectories for each agent. We show how these methodscan lead to generating emergent crowd behaviors, such as lane formation, edge effects, vortices, congestion avoidance, swirlingand modeling varying crowd density.

1 Introduction

Crowds are ubiquitous in the real world and are an impor-tant component for evacuation planning and emergency re-sponse training systems. The problem of simulating a largenumber of virtual agents and human crowds has been stud-ied in different fields including computer graphics, socialand behaviorial sciences, architecture, physics, psychology,civil and traffic engineering, and robotics. In particular,crowds are regarded as complex dynamical systems that ex-hibit distinct characteristics, such as emergent behaviors,self-organization, and pattern formation due to multi-scaleinteractions among individuals and groups.

Today’s first responders need a new generation of technol-ogy and the resources to prepare for and respond to terror-ist attacks, natural disasters, and large-scale emergencies.The next-generation simulation and experiential technolo-gies for the first responders can possibly help better pre-pare them for evacuation planning and disaster response,facilitate training experiences, and enable leaders and law-enforcement personnel to optimize their tactics, by usingwhat-if simulations based on the real situations in which theymight work and more effectively evaluating various alterna-tives. Crowd simulations with appropriate collective behav-ior models are needed for evacuation planning and trainingof first responders for unexpected events in urban environ-ments. These technologies also have diverse applications inarchitecture design, emergency evacuation, urban planning,personnel training, education and entertainment.

One of the key challenges is to simulate large-scale crowdswith tens or hundreds of thousands of agents. Such largecrowds are becoming increasingly common in cities acrossthe globe. Furthermore, some applications such as gamesor virtual environments need interactive simulation capa-

bilities, i.e. simulating at 30 frame per second (fps) ormore on current desktop systems. In addition to the overallperformance, another major challenge is generating realis-tic crowd behaviors. Despite decades of observation andstudies, collective behaviors are particularly not well un-derstood for groups with non-uniform spatial distributionand heterogeneous behavior characteristics. Such scenariosinclude pedestrian movement, evacuation flows in complexstructures, and coupled human-natural systems.

In this paper, we survey some of our recent works on ad-dressing several key computational problems of simulatinglarge-scale crowds in complex dynamic environments. Thesetechniques are necessary building blocks to compute crowdmovement, and eventually generate emergent collective be-haviors. limit our coverage here to recent methods thatare used to simulate the motion of large numbers of vir-tual agents at interactive rates. We give a brief overview ofprior work on multi-agent and crowd simulation and presenta broad classification of prior methods into various categoriesin Section 2. Next, we describe many recent approaches forlocal collision avoidance between multiple agents and globalnavigation that have been designed by our research team atthe University of North Carolina at Chapel Hill and ArmyResearch Laboratory HRED STTC. These include (a) a newlocal collision avoidance algorithm, ClearPath, between mul-tiple agents [Guy et al. 2009] in Section 3; (b) a least-effortapproach that can generate energy-efficient trajectories andemergent behaviors [Guy et al. 2010] in Section 4; (c) a novelformulation to model aggregate dynamics of dense crowds[Narain et al. 2009] in Section 5; (d) an efficient spatial con-ceptualization of avatar behaviors to effectively model per-sonalities and social protocols [Yeh et al. 2008] in Section 6;(e) an effective approach to direct and control virtual crowdsusing navigational fields [Patil et al. May 2009] in Section 7.

20th Behavior Representation in Modeling & Simulation (BRIMS) Conference 2011 - Sundance, Utah 11-BRIMS-026

177

These methods can be combined together to generate com-plex crowd behaviors that are frequently observed in urbanenvironments. Furthermore, we use the computational ca-pabilities of commodity multi-core and many-core proces-sors to exploit parallelism on modern hardware to achieveinteractive performance for tens of thousands of agents. Wealso demonstrate their application to simulating pedestriancrossings, motion of thousands of pilgrims at a mosque, tradeshows, etc.

2 Related Work

In this section, we give a brief overview of prior work relatedto multi-agent navigation and interactive crowd simulation.There is also extensive work on generating emergent behav-iors in computer animation, virtual environments and socialsciences.

Several techniques have been proposed to animate or sim-ulate large groups of autonomous agents or crowds. Mostof these methods use a rather simple representation for eachagent - this could be a circular shape in a 2D plane or a cylin-drical object in the 3D space, and compute a collision-freetrajectory for each agent. After computing the trajectoryusing a simple representation, these techniques use eitherfootstep planning or walking synthesis methods to computea human-like motion for each agent along the given trajec-tory.

At a broad level, prior methods for computing crowd move-ments can be classified into the following six categories [Guyet al. 2010]:

• Potential-based methods: These algorithms focus onmodeling agents as particles with potentials andforces[Helbing and Molnar 1995].

• Boid-like methods: These approaches are based on theseminal work of Reynolds which create simple rulesfor computing the velocities [Reynolds 1987; Reynolds1999].

• Geometric methods: These algorithms computecollision-free paths using sampling in the velocity space[van den Berg et al. 2009a] or by using optimizationmethods [Guy et al. 2009; van den Berg et al. 2009b].

• Field based methods: These algorithms tend to computefields for agents to follow [Yersin et al. 2005; Pettreet al. 2009], or generate navigation fields for differentagents based on continuum theories of flows [Treuilleet al. 2006] or fluid models [Narain et al. 2009].

• Least effort crowds: Based on the classic principle ofLeast Effort proposed by Zipf [1949], many researchershave used that formulation to model the paths ofcrowds [Still 2000; Sarmady et al. 2009]. Recently, ithas been combined with multi-agent collision avoidancealgorithms [van den Berg et al. 2009b] and used to effi-ciently and automatically generate emergent behaviorsfor a large number of agents [Guy et al. 2010].

In addition to these broad classifications, there are manyother specific approaches designed to generate crowd be-havior based on cognitive modeling and behavior [Shao andTerzopoulos 2005; Yu and Terzopoulos 2007] or sociologicalor psychological factors [Pelechano et al. 2007]. Other tech-niques include directing crowd behaviors using guidance field

specified by the user or extracted from videos and computingsmooth navigation fields [Patil et al. May 2009].

3 ClearPath – Collision Avoidance forMulti-Agent Simulations

In our work presented here, we assume that the scene con-sists of multiple heterogeneous agents, each of which is mov-ing toward an independent goal. The goal position canchange or some agents may only have intermediate goals.Furthermore, the scene consists of static and dynamic ob-stacles and it is important that each agent avoids collisionswith other agents and all the obstacles. The behavior of eachagent is governed by some extrinsic and intrinsic parametersand computed in a distributed manner for each agent inde-pendently. The overall simulation proceeds in discrete timesteps and we update the state of each agent, including itsposition and velocity during each time step based on its goalposition and the other agents or obstacles in the scene.

Our overall framework assumes that the agents are movingon a 2D plane, though our approach can be extended to han-dle agents moving in 3D space. At any time instance, eachagent has the information about the position and velocityof nearby agents. The proximity information in terms ofnearby agents can be computed using a kD-tree. Moreover,the simplest representation of each agent corresponds to acircle or a convex polygon.

Our approach is decomposed into two levels. The higher-level deals with the global path planning towards the goalusing a precomputed roadmap [LaValle 2006], and the lower-level addresses local collision avoidance and navigation us-ing ClearPath [Guy et al. 2009]. We assume that the otheragents are dynamic obstacles whose future motions are pre-dicted as linear extrapolations of their current velocities.ClearPath provides a principle to select a velocity for agentAi and implicitly assumes that the other agents Aj use sim-ilar collision avoidance reasoning. Essentially we pose thelocal collision avoidance problem for N agents as a combi-natorial optimization problem, subject to motion constraintswe impose on the system.

Any feasible solution to all of constraints, which are sep-arately formulated for each agent, will guarantee collisionavoidance. We solve this problem as a quadratic optimiza-tion function with non-convex linear constraints for eachagent. It can be shown to be NP-Hard [Guy et al. 2009] fornon-constant dimensions via reduction to quadratic integerprogramming. It has a polynomial time solution when thedimensionality of the constrains is constant – two in our case.In practice, ClearPath is more than one order of magnitudefaster than prior velocity-obstacle based methods [van denBerg et al. 2008]. Please see the video demonstration at:

http://gamma.cs.unc.edu/CA/ClearPath.mov.

4 The Least-Effort Crowd

Many researchers in different fields have noticed that the hu-man and crowd motion in real-world scenarios is governedby the principle of least effort (PLE). One of the earliestworks in this area is Zipf’s classic book on human behavior[Zipf 1949]. The basic essence of this principle is that hu-mans tend to move through the environment to their goalsusing the least amount of effort, by minimizing the time,


178

(a) Simulated crossing at Shibuya Station (b) A different view of the simulated crossing (c) A still from video of the crossing

Figure 1: The least-effort crowd simulation algorithm can automatically generate many emergent crowd behaviors at inter-active rates in the simulation of Shibuya Crossing (left, middle) that models a busy crossing at the Shibuya Station in Tokyo,Japan. (right). The trajectory for each agent is computed based on minimizing an effort function. There is a high correlationbetween the trajectories computed by the least-effort crowd simulation algorithm and the real trajectories captured by a video.

cost, congestion, or the change in velocity of their trajec-tory. The least-effort formulation has already influenced thedesign of some recent crowd modeling systems based on cel-lular automata [Still 2000; Karamouzas et al. 2009; Sarmadyet al. 2009]. However, it is hard to extend these methods tosimulate large crowds with thousands of agents.

Recently, we present an optimization-based algorithm togenerate energy-efficient trajectories for each agent [Guyet al. 2010]. Their formulation tends to follow two principlesin terms of computing the trajectories:

• Take the shortest available route to the destination.

• Attempt to move at the underlying preferred speed foreach agent.

A key issue in this formulation is the choice of the functionused to represent the “effort”. We introduce a novel metricto model the least-effort function based on bio-mechanicalprinciples in terms of the total energy expended by an in-dividual during the locomotion, measured in Joules [Guyet al. 2010]. Based on this mathematical formulation, theproblem of computing an appropriate energy-efficient trajec-tory for each agent in the crowd is reduced to an optimiza-tion problem. The resulting algorithm takes this functioninto account and computes an appropriate collision-free mo-tion for each agent that computes a path towards the goal.The performance of the overall algorithm has been tested indifferent scenarios and also compared with the trajectoriescomputed by many of the prior crowd simulation methods.Furthermore, the resulting trajectories are compared withthe prior data collected on real humans [Fruin 1971] and theauthors observed a close match. A key characteristic of thisalgorithm is that it is automatically able to generate manyemerging behaviors in various simulations. These include:

• Jamming/Bottlenecks - congestion forms at narrowpassages;

• Arching - arches form at exits;

• Lane formation - opposing flows pass through eachother;

• Wake effect - not immediately filling in space after ob-stacles;

• Uneven densities - regions form with more or less peoplethan the surrounding areas;

• Edge effects - agents move faster near the edges ofcrowds;

• Overtaking - faster individuals move past slower neigh-bors;

• Congestion avoidance - individuals tend to avoid overlydense regions if possible;

• Swirling - vortices can form in cross flows.

Furthermore, it is relatively simple to parallelize the al-gorithm on a multi-core processor and get almost linearspeedup with the number of cores. Please see the videodemonstration at:

http://gamma.cs.unc.edu/PLE/movies/PLE-shibuya.mov.

5 Hybrid Crowds

Dense crowds exhibit a low interpersonal distance and a cor-responding loss of individual freedom of motion. This ob-servation suggests that the behavior of such crowds may bemodeled efficiently at a coarse level, treating its motion asthe flow of a single aggregate system. Based on such anabstraction, we develop a novel inter-agent avoidance modelthat decouples the computational cost of local planning fromthe number of agents, allowing very large-scale crowds con-sisting of hundreds of thousands of agents to be simulated atinteractive rates [Narain et al. 2009]. A key characteristic oflarge-scale crowds is not only the number of agents, but alsothe density of the agents in terms of number of agents persquare meter. Whenever the density goes beyond 4 agentsper square-meter, which is considered very high, the crowdis treated as a whole. This includes modeling the crowd asa fluid and a continuum that responds to local influencesby assuming that the individuals in the continuum move soas to optimize their behavior to reach non-local objectives[Hughes 2002].

Our overall method for modeling very large-scale crowdscombines a Lagrangian representation [Helbing and Molnar1995; Helbing et al. 2005] of individuals with a coarser Eule-rian crowd model [Hughes 2002; Hughes 2003; Treuille et al.2006], thus capturing both the discrete motion of each agentand the macroscopic flow of the crowd. In dense crowds,the finite spatial extent occupied by humans becomes a sig-nificant factor. This effect introduces new challenges, asthe flow varies from freely compressible when the density is


179

(a) (b)

(c) (d)

Figure 2: Some examples of large, dense crowds simulated our hybrid algorithm by Narain et al. (a) 100,000 pilgrims movingthrough a campsite. (b) 80,000 people on a trade show floor. (c) 10,000 agents attempt to cross over to the opposite points ofa circle and meet in the middle, before forming a vortex to resolve the deadlock. (d) 25,000 pilgrims with heterogeneous goalsat Masjid al-Haram, a mosque built around the Kaaba, during the Hajj.

low to incompressible when the agents are close together.This characteristic is shared by many other dynamical sys-tems consisting of numerous objects of finite size, includinggranular materials, hair, and dense traffic. The new mathe-matical formulation to model the dynamics of such aggregatesystems in a principled manner can be found in [Narain et al.2009].

Results: Fig. 2 shows some of the results for dense crowdsof up to 100,000 agents closely packed in complex scenes sim-ulated at interactive rates. The authors measured the perfor-mance of their algorithm on an Intel Core i7-965 machine at3.2 GHz with 6 GB of RAM. Even with very large numbersof agents, they can achieve close to interactive performance.This method supports general scenarios with independent,heterogeneous agents and the number of unique goals hasno effect on the performance of the system. Please see thevideo demonstration at:

http://gamma.cs.unc.edu/DenseCrowds/dense-crowd.mov.

6 Composite Agents

We introduce the notion of “composite agents” to effectivelymodel different avatar behaviors for agent-based crowd sim-ulation. Each composite agent consists of a basic agent thatis associated with one or more proxy agents [Yeh et al. 2008].A proxy agent possesses the same kind of external properties

as an agent, i.e. if every agent’s external property consistsof the velocity, then the proxy agent must have its own ve-locity as well. The values of these external properties, canbe different, e.g. the proxy can possesses a velocity thatis different from anyone else. The external properties of aproxy Pj is denoted as εj . The internal state ιj of a proxyagent, however, need not be the same set of properties ofthe basic agent. We also define that Pj has access to theinternal state ιi of its parent Ai. We denote the set of allproxy agents being simulated as Proxies =

Si proxy(Ai)

This formulation allows an agent with given physical prop-erties to exercise its influence over other agents and the en-vironment. We can consider such an influence in term ofspatial extent and/or personal space of each agent, as wellas ’aura’ that some agents can impose on others nearby.Composite agents can be introduced to most agent-basedsimulation systems and used to model emergent behav-iors among individuals. These behaviors include aggres-sion, impatience, intimidation, leadership, trailblazing andapproach-avoidance conflict, etc. in complex scenes. In prac-tice, there is negligible overhead of introducing compositeagents in crowd simulation.

Fig. 3 demonstrates a variety of behaviors generated usingour novel concept of Composite Agents on top of any agent-based simulation. Please see the video demonstration at:

http://gamma.cs.unc.edu/CompAgent/CompAgent.avi


180

Figure 3: Modeling of Complex Behaviors. Top row depicts an emergency evacuation in an office building. The agentsin red are aggressive agents. They are able to carve their own way through the crowd and exit more quickly than the others.The bottom row simulates a crowd of protesters outside an embassy. Two files of policemen clear the protesters off the road.Notice that when forcing their way into the crowd, even if the actual gap between the individual policemen is enough for aprotest or to pass through, the perceived continuity of authority prevents the protesters from breaking the barricade.

7 Directing Crowds Using NavigationFields

Most existing agent-based systems assume that each agent isan independent decision making entity. Some of the meth-ods also focus on group-level behaviors and complex rulesfor decision making. The problem with these approaches isthat interactions of an agent with other agents or with theenvironment are often performed at a local level and cansometimes result in undesirable macroscopic behaviors. Dueto the complex inter-agent interactions and multi-agent col-lision avoidance, it is often difficult to generate desired crowdmovements or motion patterns that follow the local rules.

In a recent work, we address the problem of directing theflow of agents in a simulation and interactively control thesimulation at run time [Patil et al. May 2009]. Their ap-proach is mainly designed for goal-directed multi-agent sys-tems, where each agent has knowledge of the environmentand a desired goal position at each step of the simulation.The goal position for each agent can be computed from ahigher-level objective and can also dynamically change dur-ing the simulation.

This approach for directing the crowds uses discretized guid-ance fields to direct the agents. These guidance fields corre-spond to a vector field, which are used to specify the motiondirection of the agents. Based on these inputs, they com-pute a unified, goal-directed, smooth navigation field thatavoids collisions with the obstacles in the environment. Theguidance fields can be edited by a user to interactively con-trol the trajectories of the agents in an ongoing simulation,while guaranteeing that their individual objectives are at-tained. The microscopic behaviors, such as local collision

avoidance, personal space and communication between indi-vidual agents, are governed by the underlying agent-basedsimulation algorithm.

Results: This approach is general and applicable to avariety of existing agent-based methods. The usefulness ofthis approach is illustrated in the context of several simula-tion scenarios shown in Fig. 4. The user edits the simulationby specifying guidance fields, that are either drawn by theuser or extracted from a video sequence. The overall ap-proach can be useful from both artistic and data-driven per-spectives, as it allows the user to interactively model somemacroscopic phenomena and group dynamics. Please see thevideo demonstration at:

http://gamma.cs.unc.edu/DCrowd/DCrowd.avi.

8 Application to An Evacuation Scenario

Our crowd simulation system synthesizes these methods de-scribed above and finds biomechanically realistic paths foreach individual agent in the simulation. To navigate in thevirtual environment, agents walks towards their desired goalor exit while avoiding others nearby. To validate our ap-proach, we compare the results predicted by our simulationtechniques to those collected in real-world evacuation exper-iments. Figure 5(a) shows a diagram of an overhead viewof one such scenario. Figure 5(b) shows an still from oneof the runs in progress, where a participant had just passedthe exit. We analyzed data from four different exit studies.While there are minor variations between studies as shown inFigure 6, there are clear consistent trends that hold amongthem. We varied the width of the simulated exit and cal-


181

(a) Evacuation scene (b) Still image from one run of experi-ments (Kretz et al. 2006)

Figure 5: Evacuation Experiment (a) Overhead diagram of the

evacuation scenario. Participants (circles) are asked to move down

the hall (vertical lines) through the narrow passage. (b) Still image

of participants taking part of study. The walls can be adjusted to

widen or narrow the exit.

culated the flow rate of the simulated individuals past theexit. The results of this study shown in Figure 6 indicatesthat as with the human data [Kretz et al. 2006], increasingthe exit width increases the flow through the exit. More im-portantly, if we compare the flow rates at any given width,we see that the predicted flow consistently falls within therange observed in the human studies for a wide variety ofexit widths.

Figure 6: Flow Rate Comparison A comparison of the effect of

exit width (horizontal axis) on the flow rate (vertical axis) for real

humans (dashed lines) and simulated individuals (solid line). Sim-

ulated individuals have similar exit rates as the real humans for a

variety of exit widths.

9 Discussion

In this paper, we have presented a brief survey of algorithmsfor real-time behavior modeling and simulation of large-scalecrowds in dynamic scenes. We also highlighted their appli-cations on interactive crowd simulation in urban simulationsand virtual environments.

Among these methods, ClearPath is designed to take advan-tage of upcoming many-core architectures to achieve real-time collision avoidance of hundreds of thousands of dis-crete agents at interactive rates. In contrast, the ‘hybridcrowd’ approach for large-scale crowds is better suited for

highly dense scenarios, capable of simulating up to millionsof individual agents in nearly real time by exploiting theirconstrained movement. The approach to direct and controlcrowds using navigation fields is general and versatile. Itcan use any low-level collision avoidance algorithm and iscomplimentary to most current methods that compute tra-jectories for the agents. The resulting system allows the userto specify the navigation fields by either sketching paths di-rectly in the scene via an intuitive authoring interface or byimporting motion flow fields extracted automatically fromcrowd video footage, in order to generate the desired crowdbehaviors. This technique is complementary to other meth-ods and most suitable for interactive editing and hypothesistesting for ‘what-if’ scenarios.

Finally, the algorithms to simulate least-effort crowds andcomposite agents can be combined and used in any multi-agent simulation systems. In practice, these algorithms gen-erate natural-looking trajectories and crowd behaviors, aswell as individual agents that we observe. They are alsorelatively easy to parallelize on multi-core architectures.

These techniques are promising and effective for modelingand simulation of many virtual agents in urban settings forinteractive applications, such as training of first responders.Based on the techniques surveyed in this paper, we are con-tinuing to develop application systems and interfaces to en-able more efficient evacuation planning and more effectivecrowd control for emergency response. For example, we canextract the crowd flows from a live recording of crowds in apanic situation due to an unexpected event using computervision techniques. Then we use the extracted crowd flows todrive a crowd simulation through navigation fields (Section7) to “replay” the past and simulate the alternative ‘what-if’scenarios in the future with a combination of methods fromSections 3-6.

Acknowledgments: We would like to thank all our stu-dents and collaborators, whose work have been described inthis survey. These include Jur van den Berg, Jatin Chuggani,Sean Curtis, Pradeep Dubey, Abhinav Golas, Stephen Guy,Rahul Narain, and Sachin Patil. The research presented hereis supported in part by RDECOM.

References

Fruin, J. 1971. Pedestrian planning and design. Metropoli-tan Association of Urban Designers and EnvironmentalPlanners.

Guy, S., Chhugani, J., Kim, C., Satish, N., Lin, M. C.,Manocha, D., and Dubey, P. 2009. Clearpath: Highlyparallel collision avoidance for multi-agent simulation.Proc. of ACM SIGGRAPH/Eurographics Symposium onComputer Animation, 177–187.

Guy, S., Chuggani, J., Curtis, S., Dubey, P., Lin, M.,and Manocha, D. 2010. Pledestrians: A least-effort ap-proach to crowd simulation. Proc. of Eurographics/ACMSIGGRAPH Symposium on Computer Animation.

Helbing, D., and Molnar, P. 1995. Social force modelfor pedestrian dynamics. Physical Review E 51 (May),4282. Copyright (C) 2008 The American Physical Society;Please report any problems to [email protected].

Helbing, D., Buzna, L., Johansson, A., and Werner,T. 2005. Self-organized pedestrian crowd dynamics: Ex-


182

periments, simulations, and design solutions. Transporta-tion Science 39 , 1–24.

Hughes, R. L. 2002. A continuum theory for the flow ofpedestrians. Transportation Research Part B: Methodolog-ical 36 , 507–535.

Hughes, R. L. 2003. The flow of human crowds. AnnualReview of Fluid Mechanics 35 , 169–182.

Karamouzas, I., Heil, P., Beek, P., and Overmars, M.2009. A predictive collision avoidance model for pedes-trian simulation. Proc. of Motion in Games, 41–52.

Kretz, T., Grunebohm, A., and Schreckenberg, M.2006. Experimental study of pedestrian flow through abottleneck. Journal of Statistical Mechanics-Theory andExperiment .

LaValle, S. M. 2006. Planning Algorithms.Cambridge University Press (also available athttp://msl.cs.uiuc.edu/planning/).

Narain, R., Golas, A., Curtis, S., and Lin, M. C. 2009.Aggregate dynamics for dense crowd simulation. ACMTransactions on Graphics (Proc. of ACM SIGGRAPHAsia).

Patil, S., van den Berg, J., Curtis, S., Lin, M. C., andManocha, D. May 2009. Directing crowd simulations us-ing navigation fields. Tech. rep., Department of ComputerScience, University of North Carolina.

Pelechano, N., Allbeck, J. M., and Badler, N. I.2007. Controlling individual agents in high-densitycrowd simulation. Proceedings of the 2007 ACM SIG-GRAPH/Eurographics symposium on Computer anima-tion, 99–108.

Pettre, J., Ondrej, J., Olivier, A., Cretual, A., andDonikian, S. 2009. Experiment-based modeling, simula-tion and validation of interactions between virtual walk-ers. In Symposium on Computer Animation, ACM, 189–198.

Reynolds, C. W. 1987. Flocks, herds and schools: A dis-tributed behavioral model. ACM SIGGRAPH ComputerGraphics 21 , 25–34.

Reynolds, C. W. 1999. Steering behaviors for autonomouscharacters. Game Developers Conference.

Sarmady, S., Haron, F., and Hj, A. Z. 2009. Model-ing groups of pedestrians in least effort crowd movementsusing cellular automata. Proc. of 3rd Asia InternationalConference on Modeling and Simulation, 520–525.

Shao, W., and Terzopoulos, D. 2005. Autonomouspedestrians. In SCA ’05: Proceedings of the 2005 ACMSIGGRAPH/Eurographics symposium on Computer ani-mation, ACM Press, New York, NY, USA, 19–28.

Still, G. 2000. Crowd Dynamics. PhD thesis, Universityof Warwick, UK.

Treuille, A., Cooper, S., and Popovic, Z. 2006. Con-tinuum crowds. Proc. of ACM SIGGRAPH , 1160 – 1168.

van den Berg, J., Patil, S., Seawall, J., Manocha, D.,and Lin, M. C. 2008. Interactive navigation of individualagents in crowded environments. Proc. of ACM Sympo-sium on Interactive 3D Graphics and Games, 139–147.

van den Berg, J., Guy, S. J., Lin, M., and Manocha,D. 2009. Reciprocal n-body collision avoidance. In Inter-national Symposium of Robotics Research.

van den Berg, J., Seawall, J., Lin, M. C., andManocha, D. 2009. Virtualized traffic: Reconstruct-ing traffic flows from discrete spatio-temporal data. Proc.of IEEE Virtual Reality Conference, 183–190.

Yeh, H., Curtis, S., Patil, S., van den Berg, J.,Manocha, D., and Lin, M. C. 2008. Composite agents.Proc. of ACM SIGGRAPH/Eurographics Symposium onComputer Animation.

Yersin, B., Maim, J., Ciechomski, P., Schertenleib,S., and Thalmann, D. 2005. Steering a virtual crowdbased on a semantically augmented navigation graph. InVCROWDS.

Yu, Q., and Terzopoulos, D. 2007. A decision net-work framework for the behavioral animation of vir-tual humans. In Proceedings of the 2007 ACM SIG-GRAPH/Eurographics symposium on Computer anima-tion, 119–128.

Zipf, G. K. 1949. Human behavior and the principle of leasteffort. Addison-Wesley Press.

Author Biographies

MING C. LIN is currently John R. and Louise S. ParkerDistinguished Professor of Computer Science at the Univer-sity of North Carolina at Chapel Hill. Her research interestsinclude physically-based modeling, robotics and simulatedenvironments. She has published more than 220 papers inthese areas and worked closely with industry and DoD agen-cies on modeling and simulation.

DINESH MANOCHA is currently Phi DeltaTheta/Matthew Mason Distinguished Professor of Com-puter Science at the University of North Carolina at ChapelHill. His research interests include virtual environments,robotics and parallel computing. He has published morethan 300 papers in these areas and worked closely withindustry and Army Labs in terms of transitioning thesetechnologies into DoD systems.

LATIKA EIFERT is currently the Lead Principle Inves-tigator and Science and Technology Manager in EmbeddedTraining for the Ground Simulation Environment Technol-ogy Branch of the ARL-HRED-STTC. She has over 20 yearsof engineering experience including Lockheed Martin, con-sulting and now here at STTC. She has published severalpapers in training and simulation technology.

ANGEL RODRIGUEZ is currently the Branch Chief forGround Simulation Environment Technology Division at theARL-HRED-STTC. He has over 28 years in research andacquisition in government simulation and training field beingemployed with General Electric Aerospace, Naval AviationTraining Systems Division, U.S. Army Simulation Trainingand Instrumentation Command and now with STTC. Hehas published over thirteen research and acquisition papersin the area of training and simulation.


183

(a) (b) (c)

(d) (e)

(f) (g)

(h) (i)

Figure 4: (a, b, c) Motion patterns detected in a video of a crosswalk in Hong Kong with 3 frames from the original video; (d)Motion flow field; (e) Detected motion patterns; (f) Video-based motion patterns from (e) used as guidance fields; (g) Agentmotion generated by the navigation fields. (h) Sketch-based guidance fields to specify lane formation in the simulation; (i) Laneformation generated by goal-directed navigation fields.


184

interactive behavior modeling for large-scale crowd...

Documents