PaperlayerView exact revision on arXiv

Source-linked AI summary

A hierarchy of heuristic-based models of crowd dynamics

Pierre Degond, Cécile Appert-Rolland, Mehdi Moussaid, Julien Pettré, Guy Theraulaz

arXiv:1304.1927v1math-phmath.MP

TL;DR

To model crowd behavior from a heuristic pedestrian IBM, the paper derives a kinetic model and three continuum closures. The resulting hierarchy includes a game-theoretic LTE equilibrium, alongside closures suited to noiseless or noisy settings, with scope and well-posedness caveats.

  • Problem

    Reliable prediction of crowd behavior matters for public safety and crowd management, while continuum models require closure relations that can restrict their validity.

    Source: §1 · p. 2

  • Method

    The paper replaces the discrete heuristic IBM with a noisy time-continuous model, derives a kinetic model, and obtains continuum models using monokinetic, von Mises-Fisher, and LTE closures.

    Source: §1 · pp. 3–4

  • Results

    The paper presents the first kinetic and fluid models derived from this heuristic behavioral IBM, with the LTE model expressing pedestrians’ optimal directions as a Nash equilibrium.

    Source: §6 · pp. 36–37

  • Takeaways & Limitations

    The von Mises-Fisher model is presented as a compromise between physical accuracy and computational efficiency, though its hyperbolicity remains to be investigated.

    Source: §5 · p. 36

    The LTE model is restricted to local interactions and uniform vision, with no blind zone.

    Source: §5 · p. 36

Abstract

from arXiv · show

We derive a hierarchy of kinetic and macroscopic models from a noisy variant of the heuristic behavioral Individual-Based Model of Moussaid et al, PNAS 2011, where the pedestrians are supposed to have constant speeds. This IBM supposes that the pedestrians seek the best compromise between navigation towards their target and collisions avoidance. We first propose a kinetic model for the probability distribution function of the pedestrians. Then, we derive fluid models and propose three different closure relations. The first two closures assume that the velocity distribution functions are either a Dirac delta or a von Mises-Fisher distribution respectively. The third closure results from a hydrodynamic limit associated to a Local Thermodynamical Equilibrium. We develop an analogy between this equilibrium and Nash equilibia in a game theoretic framework. In each case, we discuss the features of the models and their suitability for practical use.

1 Introduction

Crowd safety and economic stakes motivate more reliable prediction, and this paper derives kinetic and continuum models from a heuristic pedestrian IBM. It presents a modeling hierarchy with three closures, while noting the assumptions and reduced validity that accompany continuum reductions.

1 Introduction

  • Crowd safety concerns and the economic stakes of managing large events motivate improved understanding and prediction of crowd behavior.The paper situates its work among individual-based, kinetic, and continuum approaches to crowd modeling.Source: §1 · p. 2
  • Kinetic models reduce the IBM to one-particle distributions, while continuum models further reduce velocity statistics to a small number of moments.The kinetic reduction ignores particle correlations.Source: §1 · pp. 2–3
  • Continuum and kinetic models can be more computationally efficient than IBMs for large crowds, but their closure relations limit their validity range.The paper argues that continuum models remain valuable for large-scale crowd analysis and prediction.Source: §1 · p. 3
  • Pedestrians are modeled as assessing encounter danger and choosing directions that balance collision avoidance with progress toward their targets.The decision is framed as selecting a Nash-equilibrium direction.Source: §1 · p. 3
  • The paper derives a kinetic model and three continuum closures from a noisy, time-continuous version of a heuristic IBM, assuming constant pedestrian speed.The kinetic model tracks the distribution over position, velocity, and target direction; the IBM-to-kinetic passage is not developed rigorously.Source: §1 · p. 3
  • The monokinetic closure is exact only without noise, the von Mises-Fisher closure accommodates nonzero noise, and the LTE closure links optimal mean velocities to Nash equilibrium.The LTE derivation requires spatially local interactions and isotropic interaction regions.Source: §1 · pp. 3–4

2 The Heuristic-Based model of pedestrian dynamics

The heuristic IBM models pedestrians as perceiving potential encounters and optimizing their direction of travel. Its perception measures feed a vision-constrained decision rule, under constant-speed assumptions that limit the model’s scope.

2.2 Perception phase

  • For a binary encounter, MD measures closest separation and DTI measures distance to the interaction point; TTI is the signed time to reach it.With several partners, the global DTI is the minimum across binary encounters; past interactions or sufficiently large MD are treated as no interaction.Source: §2.1 · pp. 4–5

2.1 Principles

  • The IBM combines encounter perception with a nonlinear optimization that balances collision avoidance against maintaining progress toward the target.The new direction is selected from tested directions in the vision cone to place the estimated endpoint closest to the target point.Source: §2.1 · pp. 4–5

2.3 Decision-making phase

  • The decision rule tests directions within a vision cone and chooses the estimated endpoint closest to the target, while assuming constant speed and omitting fixed obstacles.Because pedestrians cannot slow or stop during close encounters, the authors restrict validity to low densities.Source: §2.1 · p. 5, §2.3 · p. 9

2.4 Summary of the Heuristic-Based IBM model

  • The IBM uses fixed time steps for position updates, a choice retained to support a time-continuous limit and subsequent kinetic modeling.As the time step tends to zero, the authors say the concern about different pedestrians receiving position updates at different times disappears.Source: §2.4 · p. 10

3 Mean-field kinetic model

The paper turns a heuristic pedestrian IBM into a continuous kinetic model by smoothing its perception and directional updates, then derives a local-interaction approximation. The resulting model tracks a distribution over position, velocity direction, and target direction, with constant speed and velocity noise.

3.1 Methodology

  • Harmonic averaging approximates the IBM’s minimum-distance perception, while a continuous potential replaces discrete direction jumps; velocity noise yields smooth velocity profiles that support smooth macroscopic closures.The harmonic average gives greater weight to smaller distances, closely mimicking the original minimum operation.Source: §3.1 · p. 11

3.2 Modified time-continuous IBM

  • The time-continuous IBM bounds small interaction distances with ℓ and caps the global distance-to-interaction at L; its local potential-minimum rule may differ from the original global choice.The authors note that actual pedestrian choice between a global minimum and a nearby local one is unclear pending experiments.Source: §3.2 · pp. 12–13

3.3 Mean-field kinetic model

  • The kinetic model evolves f(x, u, a, t) over position, velocity direction, and target direction, but rigorous derivation of its interaction-coupled equation is left to future work.Without interactions, the distribution satisfies a Fokker–Planck equation; the interaction force couples back to f.Source: §3.3 · pp. 13–14

3.4 Mean-field kinetic model: discussion

  • With constant speed, directional changes balance target pursuit against collision avoidance; the force decreases a potential built from distance-to-interaction averaged over pedestrians in a vision cone.The interaction term is spatially non-local before the local approximation.Source: §3.4 · p. 15

3.5 Mean-field kinetic model with local interaction

  • Assuming large-scale variation makes f approximately constant across an interaction region, the force becomes local and uses a velocity-average kernel; with no blind zone, the kernel depends only on |v − w| and the potential no longer depends on u.The kernel can be computed analytically; the local approximation follows from retaining only leading-order terms in the small-scale expansion.Source: §3.5 · pp. 16–17

4.1 Necessity of a closure Ansatz

Taking velocity moments produces macroscopic equations, but higher moments generally prevent closure, so the distribution must be related to density and mean velocity. The paper presents three closures with different assumptions and scopes.

4.1 Necessity of a closure Ansatz

  • Higher velocity moments leave the macroscopic equations for density and mean velocity unclosed, requiring an Ansatz relating the distribution function to those quantities.A hydrodynamic justification is available only for local interactions without a blind zone; the other two closure methods apply more generally but lack that justification.Source: §4.1 · p. 20
  • The monokinetic closure assumes one velocity per target direction and is valid only without noise, yielding a second-order model with possible mass concentrations.It is analogous to pressureless gas dynamics.Source: §4.1 · p. 20
  • The VMF closure assumes a von Mises-Fisher velocity distribution to represent noise and produces a second-order model with a previously unreported momentum equation.The paper presents it as more realistic than the zero-noise monokinetic assumption.Source: §4.1 · p. 20
  • For local interactions without a blind zone, the LTE closure yields a first-order mass-conservation model whose flux is determined by the LTE and its DTI consistency equation.The LTE is determined through a fixed-point relation between the potential and DTI, and is interpreted as a Nash equilibrium balancing target progress and collision avoidance.Source: §4.1 · pp. 20–21

4.2 Monokinetic closure

The monokinetic closure replaces each target-direction-specific velocity distribution with a single velocity and derives macroscopic mass and velocity equations. Its force reflects potential reduction through interactions, while the model can develop singular concentrations when trajectories meet.

4.2.1 Monokinetic closure: derivation

  • With noise removed, the monokinetic Ansatz makes the velocity distribution a Dirac delta at U and gives an exact kinetic solution when density and velocity satisfy the macroscopic equations.The mean velocity has unit norm, a constraint preserved over time if it holds initially.Source: §4.2.1 · p. 21

4.2.2 Monokinetic closure: discussion

  • The resulting equations conserve mass and evolve velocity along flow lines under a force, with target direction entering through that force.The force couples pedestrians across target directions.Source: §4.2.2 · p. 22
  • The force tends to reduce a potential encoding target satisfaction and collision avoidance, built from averaged DTI over pedestrians in the cone of vision.Under the monokinetic assumption, the DTI uses the velocities associated with the interacting pedestrians' target directions.Source: §4.2.2 · pp. 22–23
  • The force is spatially non-local, reflecting anticipation of collision likelihood, and its local version still requires integration over target directions.When there is no blind zone, the DTI no longer depends on the target direction.Source: §4.2.2 · p. 23
  • Meeting trajectories with the same target direction can produce discontinuous velocity and density blow-up because the non-local force may be too weak to repel them.The paper motivates the VMF closure as a way to introduce internal energy to prevent this blow-up.Source: §4.2.2 · p. 23

4.3 VMF closure

The VMF closure takes the first two moments of the noisy kinetic equation and assumes velocities follow a von Mises-Fisher distribution. It yields mass and momentum equations whose noise, stress, and force terms differ from the monokinetic model.

4.3.1 VMF closure: derivation

  • Taking the first two moments of the noisy kinetic equation yields mass and momentum equations, unlike the exact monokinetic Ansatz, which needs no moment derivation.The VMF moment-based approach incorporates the noise term.Source: §4.3.1 · p. 23, §4.3.2 · p. 27
  • The closure assumes a von Mises-Fisher velocity distribution centered on the mean direction of U, with concentration β uniquely determined by |U| when |U| < 1.Small |U| corresponds to an almost isotropic distribution; values near 1 give a strongly peaked distribution.Source: §4.3.1 · pp. 24, 26

4.3.2 VMF closure: discussion

  • The VMF model’s hyperbolicity and the impact of its unusual transverse term remain for future study.The full system evolves density and mean velocity, with target directions coupled through the force.Source: §4.3.1 · p. 27, §4.3.2 · p. 28
  • The VMF model’s momentum equation adds a transverse U⊥⊗U⊥ term and scales the U⊗U term by γ∥(|U|) < 1 relative to the monokinetic closure.The transverse term arises from the VMF closure, which reflects the unit-norm constraint on microscopic velocities.Source: §4.3.2 · pp. 27–28
  • Noise produces a velocity-damping term because velocity diffusion makes the distribution more isotropic and drives the average velocity toward zero.The diffusion operator acts like the heat equation on the circle S1.Source: §4.3.2 · p. 28
  • The force is averaged over the VMF velocity distribution, and its non-local expression depends on both mass and velocity distributions.The averaged direction-to-interaction calculation also uses the VMF distribution.Source: §4.3.2 · p. 28

4.4 Hydrodynamic limit

The hydrodynamic limit assumes local interactions with no blind zone and rapid velocity changes from force and noise. It yields a density evolution law whose velocity is determined by a local equilibrium that the paper interprets as a Nash equilibrium.

4.4.1 Hydrodynamic limit: derivation

  • The limit applies to local interactions with no blind zone, and assumes force- and noise-driven velocity changes occur on very short time scales.A small parameter ε rescales the force and diffusion constants to express this timescale assumption.Source: §4.4.1 · p. 29
  • The equilibrium distribution has the form ρMDρ, where Dρ must solve a functional consistency equation; the existence and number of solution branches are left open.The derivation assumes at least one isolated branch of solutions.Source: §4.4.1 · p. 32
  • The limiting model evolves target-direction density through a single continuity equation, with mean velocity determined from the local equilibrium and density.Different target directions are coupled through the velocity computation, which depends functionally on densities across directions.Source: §4.4.1 · pp. 32–33
  • Computing the equilibrium and resulting velocity is local in space-time but may require high computational power at every discretization point.The local computation is suited to massively parallel or graphical computers.Source: §4.4.1 · p. 33, §4.4.2 · p. 33

4.4.2 Hydrodynamic limit: discussion

  • At each point and time, solving the equilibrium equation expresses pedestrians’ velocity optimization as a Nash equilibrium in which velocity is the strategy.The target direction is the players’ type, and utility is the negative of the potential.Source: §4.4.2 · pp. 33–34

5 Comparison with previous work

The models connect to existing crowd-modeling approaches through their interaction rules and flux prescriptions, while retaining features tied to the underlying pedestrian behavior. Their practical trade-offs differ: the VMF closure balances accuracy and efficiency but needs hyperbolicity analysis, whereas the third model has local-interaction and computational-cost constraints.

Time-continuous IBM

  • The time-continuous IBM combines binary interactions nonlinearly and uses a few calibration parameters, unlike linear obstacle-potential addition in the steering-potential model.Its potential resembles a cost function, but it omits acceleration costs and replaces proximity cost with a distance constraint.Source: §5 · p. 34

Mean-field kinetic model

  • The mean-field kinetic model is among the first crowd kinetic models based on detailed pedestrian-behavior analysis; gives a general framework without those interaction details.Source: §5 · p. 34

First-order models

  • The hydrodynamic model resembles first-order crowd models whose flux direction is determined implicitly by a PDE, but its direction equation also uses velocities and target velocities.Compared with, it includes velocity information as well as total density in travel-time estimation.Source: §5 · p. 35

Second-order models

  • The monokinetic and VMF closures resemble moment-closure fluid models, but constant-norm velocities motivate VMF rather than a Gaussian Maxwellian distribution.The VMF closure yields features absent from standard fluid models; the monokinetic closure is close to a zero-temperature fluid equation.Source: §5 · pp. 35–36

Comparison of the three models

  • The third, game-theoretic model captures consensus and traffic efficiency in crowded settings, but assumes local interactions and uniform vision and requires solving a costly fixed-point equation.Its local-interaction approximation is suitable at large scales when interaction regions are small relative to the scene.Source: §5 · p. 36
  • The VMF model is a compromise between physical accuracy and computational efficiency, but its usability depends on hyperbolicity; the monokinetic model is locally well-posed only until trajectories meet.At trajectory meetings, monokinetic dynamics produce unphysical mass concentrations; restoring variable speeds is proposed as a possible remedy and future work.Source: §5 · p. 36

6 Conclusion

The paper derives kinetic and fluid models from a heuristic behavioral IBM, including three distinct fluid closures. The resulting hierarchy includes a Nash-equilibrium flux model, while theoretical, numerical, and experimental assessment remains future work.

6 Conclusion

  • The paper derives a kinetic and macroscopic model hierarchy from the heuristic behavioral IBM, in which pedestrians optimize travel toward goals subject to collision avoidance.Source: §6 · p. 36
  • The first two fluid models evolve density and mean velocity using monokinetic and VMF closures, while the third uses one mass-conservation equation with a density-dependent flux.Source: §6 · pp. 36–37
  • The third model's flux relation expresses a Nash equilibrium in which pedestrians choose optimal target-directed motion amid others and noise; theoretical, numerical, and experimental study remains future work.Source: §6 · p. 37

Related papers

Loading 1304.1927v1…