2015-stuvel-collision-perception-poster-ictopen.pdf

(818 KB) Pobierz
Perception of Collisions between Virtual Characters
Sybren A. Stüvel
Arjan Egges
A. Frank van der Stappen
Virtual Human Technology Lab
Universiteit Utrecht
http://vhtlab.nl/
Abstract
With the growth in available computing power, we see increasingly
crowded virtual environments.
In densely crowded situations
collisions
are
likely
to occur. The choice in collision detection technique can impact
the maximum density obtainable with a real-time crowd, and the perceived
realism of the crowd.
We present an investigation into the accuracy of human observers with
regard to the recognition of collisions between virtual characters.
●
●
Variables
Character angle
α ∈
{45, 90, 135, 180} degrees.
The severity
S
of the (near) collision labelled as LOW, MEDIUM1, MEDIUM2 or
HIGH, and expressed either as I
V
when colliding or D
m
otherwise.
|L| = D
m
User Study
In our main experiment, we showed the participants 32 videos depicting
16 colliding and 16 non-colliding situations.
In the additional experiment we used static images, and the following variables:
●
Mesh-mesh distance Dm
∈
[−0.10, 0.20] metres
●
λ
∈
[0, ∞) measures the length of the visible (i.e. not occluded by the front
character) part of L, measured in metres.
Analysis
Animated characters:
Colliding:
the most important factor was
α
Non-colliding:
the most important factor was
D
m
In our second experiment, we showed each participant 50 static images
depicting 25 colliding and 25 non-colliding situations.
Static characters:
Results & Conclusions
Participants show a
bias
towards answering “not colliding”.
Asymmetrical:
slightly colliding cases hardest to recognise; a penetration
depth of 3 cm shows the lowest accuracy.
Colliding
→
angle
between characters most important;
Non-colliding
→
distance
between characters most important.
Participants were slightly
more sensitive to collisions in the upper body
than the lower body.
For faster collision detection of humanoid characters that match our
perception,
simpliied shapes should use bounded volumes,
rather than
the commonly used bounding volumes. By ensuring a Hausdorf distance
of at most 1.5 cm. the total penetration of two such meshes would be at
most 3 cm and fall within the interval of minimal average accuracy.
Conirms the asymmetric response to (non-)
collision severity and higher
sensitivity to upper body collisions
Future Work
We are interested to see the efects of various factors on the perception
of collisions:
Diferent
shadow
rendering techniques.
●
Background
texture & static objects, such as bushes, buildings etc.
●
Moving objects
in the background, such as other crowd agents.
●
Diferent shapes,
to see whether the observed efects are speciic
to the human shape.
●
Collision
response
animations.
●
Collision
avoidance
animations, for example by slightly moving
hands or feet without changing global position & heading.
●
Scan me for a digital copy of this poster.

Plik z chomika:

Inne pliki z tego folderu:

Inne foldery tego chomika:

Zgłoś jeśli naruszono regulamin