Practical Models of Human Visual Perception for Simulation

Practical Models of Human Visual Perception for Simulation

Chris DarkenAssoc. Prof., Computer Science

MOVES 11th Annual Research and Education SummitJuly 12, 2011

831-656-7582http://movesinstitute.org

Outline

• Why we care about perception• Quick tour of recent work applied to:

– Army’s Combat XXI (joint with TRAC)– ONR’s BASE-IT prototype

• Suggested future work

1. Facilitating Training

2. Real World Perceptual Tasks

3. Human Behavior Simulation

Outline: Three Methods

• Concealed routes• Pre-cached approximate LOS• Hierarchical threat maps

Background• Two types of visual perception deficits for synthetic entities in

combat sims– Complete lack of terrain info visible to the user– Information provided, but not aggregated properly

• Significance of visual perception deficits– Hunting and avoiding threats in combat sims requires adequate visual

perception representation• This is especially true for analytic simulations of situational awareness

enhancing gear– Failure to fix deficits makes accurate behavior either

• Impossible• Too computationally inefficient for practical use• Too difficult for the typical scenario developer to implement

Outline: Three Methods

• Concealed routes• Pre-cached approximate LOS• Hierarchical threat maps

Concealed Routes

• Goal is to provide an automated way for entities to take terrain into account when planning movement.

• Conceptually a location that has LOS to a large number of other locations can be said to have “bad” concealment.

Good Concealment

Bad Concealment

Line of Sight Density Results

No LOS density cost moves over open terrain

High LOS Densitycost tended to follow rivers and ridges

Artifact caused by edge of network.

Outline: Three Methods

• Concealed routes• Pre-cached approximate LOS• Hierarchical threat maps

Point Based Estimated LOS Accuracy• Used coarse grid of nodes to approximate true LOS• Tested to compare accuracy of estimated LOS to

actual LOS.– 16.4 km2 box in Camp Pendleton, 50m node spacing, about

3% covered with buildings.• False positives 0.1% / False negatives 5.0%• Processing about 35% faster.

– 1.7 km2 box in flat fictitious terrain, 20m, about 40% covered with buildings.

• False positives 1.0% / False negatives 2.0%• Processing about 25 times faster.

Point Based Pre-calculated LOSActual LOS Point Based LOS

Outline: Four Methods

• Concealed routes• Pre-cached approximate LOS• Hierarchical threat maps

Hierarchical Threat Maps• Threat map: probability distribution providing a subjective

estimate of where threats are• Variant of ACQUIRE based on precomputed target size and

contrast values is used• Requires lots of computation!• Visibility data at left; threat map at right

Scaling Up

16

2008Fictitious terrain

1,700 m2

2009Partial Range 200, 29 Palms

13,000 m2

2010Kilo 2, Pendleton

155,000 m2

Solution: Multiple Levels of Resolution

• How to use – Coarse resolution far from entity– Fine resolution near

• Why to use– Saves computation– Intuitively accurate

From Clingman et. al., “Practical Java Game Programming”

Fireteam 1’s Threat Model

Fireteam 2’s Threat Model

Results

• Number of nodes in hierarchical vs. original model (left)

• Model update time for both models (right)

Conclusions

• Representations for concealed movement planning, visual search, and threat location have been presented

• Precomputation and muti-resolution spatial data structures make the techniques practical

• All have been prototyped either in Combat XXI (TRAC) or the BASE-IT (ONR) prototype sims

• Much work remains to be done, both towards validation and obvious extensions of the models

22

From 2004

Problem: Nearly Blind Agents

It Seems That We’re Always Missing Something…

• Cover– Lacked obstacle info

• Target detection– Lacked exposed surface and contrast info

• Threat Location/Visual search– Lacked both above plus shooting position info

…So Why Not Give Just Agents Everything We Give a Human User?

Questions?