Level of Detail Techniques for Crowd Simulation

1
Level of Detail Techniques for Crowd Simulation

• Stephen J. Guy

2
Why Realtime Crowds?

• Movie-like games
• Epic feeling
• Large battles
• Immersive Environments
• Simulations
• Training
• Entertainment

3
Overview

• Creating a Crowd Simulation
• Rendering
• Animation
• Simulation
• Level of Detail Techniques for Simulation
• Implementation
• Results
• Analysis
• The Future

4
Crowd Sim 101

• What is a Crowd Simulation
• Crowd simulation is the process of simulating
  the movement of a large number of objects or
  characters Wikipedia
• Agent The subject of the simulation
• What we really need to know
• Where is each agent?
• What is their state?
• What do they look like right now?
• How do we draw that quickly?

5
Rendering

• Nieve approach
• OpenGL - Render each agent as full 3D geometry
• SLOW
• Better
• Visibility Culling
• Level of Detail
• Image Based Rendering
• Imposters

6
Imposters

• Replace Far off agents with an oriented texture
• Several Issues
• Popping
• Uniformity
• Lighting
• Shadows
• Many issues addressed in recent works

7
Animation

• Naive
• Generate a walk cycle for each agent
• Chose appropriate model each time step
• Slow or speed animation based on velocity
• Better
• Use an FSM and several animations
• Motion Clips

8
Simulation

• Tells us where each agent should be each timestep
• Simulation Methods
• Continuous Methods Crowd as flow
• Cellular Automata Crowd as grid states
• N-body Crowd as force models
• Agent-based Crowd as lots of individuals

9
Cellular Automaton (CA)

• Divide Space into grids
• Assign each grid cell a value
• Devise rules for updating grid cells based on
  values of that cell and its neighbors
• Watch interesting behavior emerge

10
Agent Dynamics Based

• Simulate each agent individually takeing into
  account neighbors, goals, and obstacles as
• Forces (or)
• Constraints
• Additional forces may model
• Social forces
• Sociological factors
• Psychological models
• Apply combined forces to each agent, each time
  step

11
3 Example Simulations

• My Toy Simulator
• OpenSteer
• Craig Reynolds, 2004
• AERO
• Sud et al, 2007
• AERO is the most full featured, but all three
  focus on simulation (not rendering or animation)

12
Toy Simulator

• Model based on agents and forces
• Roadmap is generated in empty space
• Agents are given path force based on A search
  for goal
• No bells or whistles

13
OpenSteer

• Extension of flocking motion to general steering
  behavior
• Same basic idea, but
• Very fancy neighbor search
• Free and Open Source

14
AERO

• Adaptive Elastic ROadmaps (AERO)
• Introduces deformable paths for agents to follow
  (while still avoiding moving obstacles)
• Some agents are aggressive

15
Simulation Level of Detail (LOD)

• LOD is traditional a rendering technique
• Replace complex geometry with simple one when far
  away
• Adapt this idea for rendering
• Replace individual agents with groups of agents
  where possible

16
Overall Algorithm

• Loop
• Group agents into clusters
• Create über-agents from cluster
• Simulate über-agents (using existing techniques)
• Propagate changes from über-agents back to real
  agents
• Key points
• Allows us to take advantage of existing crowd
  simulations
• Similar, nearby agents will move as a group

17
Clustering

• How to group agents?
• Cluster nearby, similar agents so they can be
  treated as one über-agent
• Requirements for a cluster
• Heading difference between 2 agents can differ by
  at most v radians
• The ratio of speeds between 2 agents can be at
  most s
• A cluster can be at most g units across in length
• 2 or more agents can not have differing immediate
  goals
• A cluster may not intersect any obstacle

18
Clustering (cont)

• Cluster using a KD-tree
• Decide on criteria which clusters must follow
• Add all agents to one cluster, Spit()
• Split
• If cluster violates any above criteria
• Create 2 new clusters
• Call Split() on new cluster

19
Clustering Options

• K-means, K-d tree, R-tree, EM, etc
• K-means clustering had odd behavior

• KD-tree allows local refinement

20
Splitting Rules

• Manner of splitting can effect shape of cluster
• For a good aspect ratio
• Chose the longest axis to split on
• Place a perpendicular plane halfway along the
  axis
• Move the plane so at least one agent is on each
  side
• Agents not separated by a plane are in the same
  cluster

21
Clustering Costs

• Time
• Clustering costs grow O(n log n)
• Possible to reduce with incremental update
• Split is fine
• Merge is difficult!
• Fidelity
• Can limit approximation by capping num agents in
  a cluster
• Can be used to adapt to framerate

22
Results

• Approximation Level vs. Time Toy Demo

23
Results Toy Sim.

• Simulation Time vs. Num agents

24
Results OpenSteer

• Simulation Time vs. Num agents

25
Results - AREO

• Simulation Time vs. Num agents

26
Results Comparison
Bold is better

• With few agents clustering overhead was
  noticeable
• With several hundred agents, clustering was a
  clear win
• OpenSteer showed the least gain in performance

27
Analysis

• Simulation Level of Detail (LOD) can result in
  significant speedup
• Relatively large speed gains are possible with
  small clusters if they dont have to be split
  often
• Results were best with complex paths and heavy
  per agent costs
• As implemented, LOD was less helpful when added
  on top of other acceleration techniques

28
Future Work

• Verification
• How do we know LOD wasnt too aggressive?
• By eye?
• User studies?
• Numerically comparing paths?
• Average time to reach goal?
• Rendering Accelerations
• Simulation LOD and Rendering LOD are somewhat
  similar
• Is there some synergy that can be exploited?

29
Far Future

• Agent Model
• Intelligent Agents
• AI/Machine Learning
• Systems
• High Performance
• Mobile Crowd Sims
• Immersive Displays

You here!
30
Acknowledgment
!

• Thanks to
• Ming Lin
• GAMMA group
• My Committee members
• Funding agencies

31
Questions

• ?

32
Begin Extra Slides

• My Imposter Implementation
• Geoposters
• Flocking
• Multi-agent Navigation Graphs
• K-means vs KD-tree

33
My Imposter Attempt

• Worked, but not pretty

34
Geoposters
35
Flocking

• Seminal work in multi-agent movement
• Assign simple force to each agent
• Used in
• Lion King
• Batman Returns

36
MultiAgent Navigatoin Graph

• MANG
• One Path to Guide them All

37
Clustering Options

• K-means, K-d tree, R-tree, EM, etc
• K-means clustering had odd behavior

• KD-tree allows local refinement