Animat Vision: Active Vision in Artificial Animals

1
Animat Vision Active Vision in Artificial
Animals

• by Demetri Terzopoulosand Tamer F. Rabie

2
Animat Vision

• Whats an animat? - computational models of real
  animals situated in their natural habitats
• Animate vision - a paradigm which prescribes the
  use of artificial animals as autonomous virtual
  robots for active vision research

3
Fish Animat

• Challenge - to synthesize an active vision system
  for the fish animat, based solely on virtual
  retinal image analysis.
• Binocular perspective projection of the 3D world
  onto the animats 2D retinas.
• Use fish because they have simple goals, can swim
  in environment, rely on vision and recognition.

4
Hardware vs. Software Approach

• Hardware approach
• - Cant model the complexity of natural animals
• - Expensive
• Software approach
• - Can slow down the cosmic clock
• - The quantitative photometric, geometric, and
  dynamic information needed to render the virtual
  world is available explicitly

5
Previous Related Work

• A point marker on a 2D grid world
• 2D cockroaches
• Kinematic dog
• Animats using perceptual oracles

6
Qualities of Fish Animat

• Motor system
• Perception system
• Behavior system
• Form and Appearance

7
Motor System

• Comprises the fish biomechanical model, including
  muscle actuators and a set of motor controllers
  (MCs)

8
Perception System

• Model limitations as well as abilities
• Perceptual attention mechanism - allows animat to
  act in a task-specific way
• Perceptual oracles vs. animat vision

9
Behavior System

• Mediates between the perception system and the
  motor system of the fish

10
Form and Appearance
11
Eyes and Retinal Imaging
12
Active Vision System

• Must stabilize in environment
• Must foveate on target

13
Locating a Target

• Use Color Histogram Intersection
• Most obvious algorithm is to compare the color
  distribution of the target with color
  distributions found on the retinal image

14
Problem with obvious solution

• Only works if scale of target is similar to the
  scale of the image.
• Works poorly if object is far away
• Works poorly if object is semi-occluded.

15
Locating Targets Second Try

• Iterate over scaled versions of the image and
  take the average.
• Good Generally converges after 2-4 iterations.
• Bad Leads to false alarms if model is overly
  scaled.

16
Locating Targets Third Try

• Like before, but use a weighted average to place
  more importance on colors that are specific to
  the model.
• In their experiments, usually converged to Pgt0.8
  or Plt0.2 within a few iterations.

17
Navigation

• Once targets are located and can be tracked,
  navigation is trivial.
• When left-right vision angles deviate by more
  than 30 degrees from center, tell body to turn
  left/right.
• When up-down vision angles deviate by more than 5
  degrees, tell body to push up/down.

18
Pursuing Targets in Motion

• How does this fish perform in pursuit of another
  virtual fish?
• Ran an experiment and plotted gaze angles.
• Performed well, was not distracted by fake
  targets.

19
Plot of Pursuit
20
Picture from Pursuit 1/2
21
Picture from Pursuit 2/2
22
Conclusions, Looking Forward...

• Achieved goal of implementing a software-based
  artificial life simulation.
• In the future, would like to develop a better
  active vision algorithm more suited to real fish.
• Model can be made realistic enough to use
  resulting data to form theories about animals and
  robotic situated agents.