1. Obstacle detection using v-disparity image
Based on: Global Correlation Based Ground Plane Estimation Using V-Disparity Image, By: J. Zhao, J. Katupitiya and J. Ward, 2007
Obstacle Detection with Stereo Vision for Off-Road Vehicle Navigation, By: A. Broggi, C. Caraffi, R. I. Fedriga, P. Grisleri, 2005
Real Time Obstacle Detection in Stereovision on Non Flat Road Geometry Through ”V-disparity” Representation, By: R. Labayrade, D. Aubert, J. Tarel, 2002
Presented by: Ali Agha
March 25, 2009

2. Motivation Obstacle avoidance using stereo vision Points with Z > 0
The pitch angle between the cameras and the road surface will change. Therefore, we need to compute the pitch angle and disparity of ground pixels dynamically
due to static and
dynamic factors
thus the disparity of ground pixels is
changing from time to time.

3. Related work
Disparity map  3D point cloud  Ground plane extraction using plane fitting  obstacle detection
Yu et al (2005)
Disparity map  Optical flow and ground information  obstacle detection
Mascarenhas (2008)
Giachetti et al (1998)
Disparity map  V-disparity image and ground information  obstacle detection
Labayrade (2002)
Broggi (2005)
Zhao (2007)

4. V-Disparity image Disparity map IΔ has been computed IvΔ is built by
accumulating the
pixels of same
disparity in IΔ along
the v axis u d
Recently, V-disparity image has become popular for
ground plane estimation [1][9][10][11]. In this image, the
abscissa axis (w) plots the offset for which the correlation
has been computed; the ordinates axis (v) plots the image
row number; the intensity value is settled proportional to
the measured correlation, or the number of pixels having
the corresponding disparity (w) in a certain row (v). Each
planar surface in the field of view is mapped in a segment
in the V-disparity image [10]. Vertical surfaces in the 3D
world are mapped into vertical line segments, while ground
plane in the 3D world are mapped into slanted line segment.
This line segment, called ground correlation line in this
study, contains the information about the cameras pitch
angle at the time of acquisition (mixed with the terrain
slope information). v v

5. The image of a plane in v-disparity image
pin-hole camera model
Ground correlation line 5

6. Ideal and real road and obstacle
Recently, V-disparity image has become popular for
ground plane estimation [1][9][10][11]. In this image, the
abscissa axis (w) plots the offset for which the correlation
has been computed; the ordinates axis (v) plots the image
row number; the intensity value is settled proportional to
the measured correlation, or the number of pixels having
the corresponding disparity (w) in a certain row (v). Each
planar surface in the field of view is mapped in a segment
in the V-disparity image [10]. Vertical surfaces in the 3D
world are mapped into vertical line segments, while ground
plane in the 3D world are mapped into slanted line segment.
This line segment, called ground correlation line in this
study, contains the information about the cameras pitch
angle at the time of acquisition (mixed with the terrain
slope information).

7. Hough Transform Labayrade uses Hough transform to extract the ground
correlation line

8. free space on the road
Once the longitudinal profile of the road has been extracted,
the disparity values on the road surface are known
for each scanning line (it is exactly the disparity value corresponding
to the longitudinal profile curve extracted for
the scanning line in question). Let Δi denote this value.
Thus, it is straightforward to extract the obstacles from
the disparity map IΔ already computed: for a given scanning
line, pixels whose disparity is equal to Δi are located
on the road surface; other pixels belong to potential obstacles.
On Fig. 4, note that the vehicle located at almost 80
meters is well perceived as a potential obstacle.
Once the obstacle areas have been computed, the free
space on the road surface can be extracted (see also [6][7]),
using a growing area algorithm (see Fig. 6 Down). The
part of the image corresponding to the free space area can
be used, for example, for extracting the lane-markings in
more suitable conditions.

9. Robust ground correlation extraction
Broggi et al. experimentally, found that the ground correlation line during a pitch variation oscillates, parallel to itself.
Zhao et al. investigated this characteristic mathematically and gave the condition for this characteristic to be valid.
The behavior of the ground correlation line during a
pitch variation is to oscillate, parallel to itself. Experimentally,
we found out that the cameras height variation due
to oscillation has neglectable effects. Accumulating the Vdisparity
image values along each of the candidate lines, it
is possible to estimate the ground correlation line (choosing
the line that accumulated the greatest value), and then the
pitch of the cameras at the time of acquisition.

10. Condition derived by Zhao et al
For the ground correlation lines to be parallel with each other at different pitch angles
At different pitch angles shown in Fig. 3, the slope
of ground plane in V-disparity image will be different.
Suppose gs is the slope of ls using static calibration data
when = s, gmax is the slope of lmax when = max,
and gmin is the slope of lmin when = min as shown in
Fig. 4.
Thus for a given 4w and certain pitch oscillation, we need
to reduce s. In [10], the tilt angle is 8.5. In [1], the tilt
angle is 7.73.
Θs= is 7.73.

11. Result of this condition
If this condition is satisfied.

12. GLOBAL CORRELATION METHOD
Exploiting this characteristic  Fast and Robust method (even in lack of distinct features)
By accumulating the matching cost (intensity of V-disparity) along each of the candidate lines and choose the one with least (most) accumulated matching cost as ground correlation line.

13. Verifying equations and assumptions (Road has distinct features)
Implementing the Labayrade’s work
Disparity map
V-D
Hough

14. Verifying equations and assumptions (Road has distinct features)
In the sequence of images
the position of ground correlation lines is ranging from 25 to 28.
The slop(g) of ground
correlation lines
ranges from 5.5 to
After Hough transform, we can get the position(4w)
and slope(g) of the ground correlation lines for these image
pairs. The result is shown in Fig. 9. The lower plot of this
figure shows the position(4w) of ground correlation lines
for different image pairs, which is ranging from 25 to 28.
This plot tells us there do exist pitch variation in different
conditions. The upper plot shows the slop(g) of ground
correlation lines, which ranges from 5.5 to From
this Figure we can see that the bigger the slope(g) is, the
greater 4w. This confirms the match between Eq. 4 and
Fig. 4 in section II. Also, the upper plot shows that the
change of slope is very small, the difference between the
greatest slope and smallest slope is merely , which
is 1.17%. Thus the assumption that the ground correlation
lines under different pitch angles are parallel, is valid.

15. Results Applying the method on the same images
Calculate the matching cost
(associated with each candidate ground correlation lines)
matching costs associated with candidate
lines
Same as Hough
transform
The matching cost of each candidate line is accumulated from the bottom of the image to the level Of horizon
GCL

16. Without distinct features
Disparity map
V-D
Hough

17. Without distinct features - Comparison
Hough transform and global correlation method
GCL
Hough

18. Obstacle detection

19. Conclusion
This accuracy is dependant of the image quality and whether or not the ground pixel dominate this area of the image.
It seems an appropriate method for detecting obstacles such as vehicles in road in night using structured light (as the headlight of automobile)

20. Thank you Questions??