Computational Techniques on the Perception of Visual Motion

1
Computational Techniques on the Perception of
Visual Motion
2
Project Goals

• Investigate the perception of visual motion and
  develop tools for its evaluation and
  understanding
• Apply it in the study of drivers visual
  perception abilities, such as object motion
  segmentation and perception of heading (LPUM)
• Contribute to safer driving environments and
  behaviors (LAT)

3
Problem Description

• Different scenes generate different apparent
  motion patterns
• Motion perception is not the same even with the
  same travelling speed

4
Problem Description

• How to characterize and quantify apparent motion?
• How to relate it with perceived visual motion?

5
Related Work

• Studies on the perception of visual motion are
  mostly done at theoretical level either
  neurological, psychophysical or computational,
  but with very few relationships among them
• Autonomous navigation projects have totally
  different purposes and methodologies unsuitable
  for the task at hand

6
Concepts

• Optical flow is the variation of brightness
  patterns in an image sequence
• It it a measurement of perceived visual changes
  or apparent motion in the scene
• It is an approximation to image motion, but
  suffers from the aperture problem and may present
  photometric as well as aliasing effects
• It can be calculated by differentiation in space
  and time

7
Technology

• Standard video equipment
• less than 800x600 spatial resolution
• maximum of 25 images per second
• Pipelined real-time video processing system
• processing is performed at video frame rate
• completely programmable operations
• Sun SPARC Workstation
• Solaris Operating System (Unix)
• C/C language libraries and programs

8
Technology (cont.)
Video Camera
Sun Workstation
Computer Monitor
Video Monitor
VME Rack with MaxVideo 200
Video Tape Player/Recorder
9
Real-time video processing system

• Datacube MaxVideo 200

10
Procedures

• Spatio-temporal image diferentiation
• Causal IIR temporal filter
• Velocity computation by local least squares fit
  on each pixel
• Dense characterization of optical flow
• High computacional cost for the system

11
Current Status

• Off-line differed version concluded and used in
  experimental psychology, in studies with
  simulated visual motion
• Implementation of real-time image diferentiation
  concluded
• Computation of velocity must be performed by the
  central processor
• Embedding in a multi-scale coarse to fine strategy

12
Future Work

• Develop motion based segmentation algorithm
• Extract self-motion information to confront with
  perceived heading direction
• Portuguese-Spanish collaboration in the field of
  road traffic
• Combine optical flow techniques with drivers eye
  tracking information in real traffic scenes