M. Kardouchi & E. Hervet Universit de Moncton Moncton, NB, - - PowerPoint PPT Presentation

▶

May 03, 2023 232 likes •405 views

M. Kardouchi & E. Hervet Universit de Moncton Moncton, NB, Canada 30-mai-03 1 Introduction Approach Feature points extraction points extraction and matching and matching Feature Dense Dense vector field vector field

SLIDE 1

30-mai-03 1

M. Kardouchi & E. Hervet

Université de Moncton Moncton, NB, Canada

SLIDE 2

30-mai-03 2

Introduction Approach

Feature Feature points extraction points extraction and matching and matching Dense Dense vector field vector field Delaunay triangulation Delaunay triangulation Adaptive Adaptive interpolation interpolation

Parallelization Results and Conclusion

SLIDE 3

30-mai-03 3

Disparities between stereoscopic

Disparities between stereoscopic images images → Displacement between 2 image points

Estimation of dense

Estimation of dense correspondences correspondences → Block matching, Optical flow methods Pb: disparity discontinuities → Feature points matching and Delaunay triangulation

SLIDE 4

30-mai-03 4

Extraction of feature points with Harris Detector

Left and right images Extracted feature points

Harris detector’s threshold : 800-1200 corner points

from a 720 x 480-pixel image

SLIDE 5

30-mai-03 5

→

Maximizing a cross-correlation measure 600-900 acceptable matches for 800-1200 corners points Correspondence vectors between left and right images Problem : small number of mismatches

SLIDE 6

30-mai-03 6

Eliminating mismatches

→ Three steps

1. Delaunay triangulation

SLIDE 7

30-mai-03 7

2. Identify neighbours of each correspondence point p

Neighborhood of correspondence point p

( )

p η

SLIDE 8

30-mai-03 8

3. A correspondence point p is considered unreliable if:

p p

d d ρ − >

d

( )

p η

ρ

disparity vector at p average correspondence vector over adjustable threshold.

SLIDE 9

30-mai-03 9

Correspondences vectors before and after post-processing ~ 16% of correspondences removed

SLIDE 10

30-mai-03 10

Triangulation

Triangulation and correspondence vectors and correspondence vectors: dense : dense vector field vector field Assumption Assumption: in spatial : in spatial coordinates coordinates, , disparity disparity varies varies linearly linearly within each within each triangle. triangle. ( )

1

x p q r

d i j d id jd = − − + +

d

p

q

r

x Triangulation Disparity vector

d

SLIDE 11

30-mai-03 11

Disparity field obtained with linear interpolation Advantage: continuous dense disparity field Limit: disparity discontinuities

SLIDE 12

30-mai-03 12

Adaptive interpolation

Euclidian distance disparity / 3 vertices
Weight of vertices according to prediction error

Disparity field obtained with adaptive interpolation

SLIDE 13

30-mai-03 13

To speed up the adaptive interpolation

→ video flows processing in real time → object tracking

Theoretically a simple linear system

SLIDE 14

30-mai-03 14

In practice: computation balancing of triangles
First sort triangles according to their number of points
Then processor #1 is given t or t+1 triangles to process:
Repeat until all triangles are processed

SLIDE 15

30-mai-03 15

Cluster: 1 server + 16 nodes
server: Intel PIV processor 2GHz, 512 Mb RAM
nodes: Intel PIII 500 MHz, 128 Mb RAM
Ethernet cards 100 Mbits/s
Operating System: Linux Mandrake
Parallel development environment:

MPI (Message Passing Interface)

SLIDE 16

30-mai-03 16

4 proc.: gain 3.45 → Standard stereoscopic video: 1 image out of 3 8 proc.: gain 6.18 → video: 1 image out of 2 1 node: 414 ms

SLIDE 17

30-mai-03 17