Presentacin de PowerPoint

1
PhD Program in Information Technologies
Visual Perception
Description Obtention of 3D Information. Study
of the problem of triangulation, camera
calibration and stereovision. Passive and active
vision. Epipolar geometry and bidimensional
transformations. Coordinator Dr. Rafael
Garcia Professors Dr. Rafael Garcia Rafa, Dr.
Joaquim Salvi Quim, Josep Forest Pep. Term
March April Day Time Friday from 11 to 13
h. Place Seminari EIA

2
Contents of the Course
1. Introduction to visual perception (2 hours)
Human vision. Image interpretation brain vs.
computer. Phases of image processing. Quim CCD
sensors. Type of cameras matricial, linear, 1
CCD, 3CDD, Analog, Digital. Rafa 2. Camera
modelling and calibration (2 hours) Quim Camera
modelling, camera calibration intrinsic and
extrinsic parameters, stereo vision, epipolar
geometry, fundamental matrix. Example robot
localization and 3D mapping. 3. Motion
estimation. (4 hours) Rafa Trinocular
stereovision. Deriving homographies from the
projection matrix. Robust estimators.
Aplications motion estimation through
mosaicking. Derivation of extrinsic parameters.
3
Contents of the Course
4. The correspondence problem. (2 hours)
Rafa Detection of interest points. Finding
correspondences. Similarity measurements. Aplying
epipolar geometry. 5. 3D reconstruction using
laser range finders. (2 hours) Pep Laser beam
calibration. Subpixel slit detection. Scanning.
3D reconstruction. Examples. 6. Structured
light (2 hours) Quim Pattern projection. Pattern
coding. Time multiplexing. Spatial neighborhood.
Direct codification. Designing and Implementing
an optimal pattern. Practical issues
Modelization and calibration of a computer vision
system and reconstruction of 3D objects.
4
Schedule of the course
March 2004
April 2004
May 2004
Lesson Days
Practical Issues presentation Second week of June
5
Introduction to Visual Perception

• Human Vision
• Identify objects
• Determine the shape
• Locate its 3D position.

Image acquisition
Image interpretation
6
The Human Eye ?

• Eye shape
• Cornea Transparent surface.
• Sclera Outer cover composed of a fibrous coat
  that surrounds the choroid.
• Choroid a layer of blood capillaries.
• Retina layer inside the choroid composed of two
  types of receptors (rods and cones) and a netword
  of nerves.
• Optic nerve Retinal nerves leave the eye to the
  brain trough the optic nerve bundle.

• Image enhancement
• Cornea Transparent surface.
• Lens Focuses the light to the retina surface to
  perform proper focus of near and distant objects.
• Iris Acts as a diaphragm to control the amount
  of light entering the eye.

7
How an eye is working ?

• Image acquisition
• Retina Composed of
• 100 M. Rods Long slender receptors.
• Sensitive at low levels of light.
• 6.5 M. Cones. Shorter and thicker receptors.
• Sensitive at high levels of light.
• Greatest presence at the Fovea region (sharpest
  vision).
• Three types of cones with different wavelength
  absorption with peaks in the blue, green and red
  light spectrum
• Light stimulus activate a rod or cone producing
  a nerve impulse which is transmitted through the
  optic nerve.

More information at http//www.vision.ca/eye/lobb
y.html
8
Computer Vision Object Recognition. Object
Localisation. Advantage Automatisation. Constrai
nt Difficult to transmit the human intelligence
and skills to a computer. Applications Shape
Inspection for quality control Rapid
Prototyping Computer assisted surgery Film making
effects Object picking Robot Navigation
Image acquisition
?
?
Image interpretation
9
3D Information
System selection
Modelling
Calibration
Correspondence
Get 3D Cloud
Data Fusion
10
System Selection

• Combination of computational and optical
  techniques aimed at estimating or making explicit
  geometric (3D shape) properties of objects or
  scenes from their digital images.
• stereovision
• pattern projection
• laser scanning
• shape from X (motion, texture, shading, focus,
  zoom)
• Computation for all or some pixels of the
  distance between a known reference frame and the
  scene point that is imaged in those pixels. The
  output is a range image (depth map) or a cloud of
  points (xi, yi, zi), i1..N.
• The fusion of several range images or point
  clouds corresponding to partially different views
  of an object may yield its full 3D digitization.

11
Main processes in 3D digitization
solid(triangles)
graphicsurface
N 3D point clouds
Range sensing
Geometric fusion
Objectmodeling
object
solid (splines)
System Selection
best nextview
Sensorplanning

• Stereovision
• Pattern projection
• Laser scanning
• Shape from X
• (motion, texture, shading, focus, zoom)

Texturemapping
coloured solid
12
Geometric fusion
24 aligned 3D scans ready for merging
set of six 3D scans acquired from different
viewpoints and their alignment (center)
24 meshes merged into a surface triangulation.