Multiple Plane Detection Method from Range Data of Digital Imaging System for Moving Robot Applications - Visual Computing: Scientific Visualization and Imaging Systems - page 200

In step ´, we find the maximum peak in Hough parameter space. The peak
represents the normal direction of the biggest plane in the 3D range data. In step ˆ,
if the voting count of the peak is lower than TH1, we determine that there is no
plane and the plane segmentation algorithm is ended. In ˜, we find the plane data
in the 3D range data by using Eq. ( 11.9 ), which is derived from the plane equation.
If the value is lower than a given threshold value TH2, the data is saved into the ith
plane data and it is removed for detection of the (i + 1)th plane. Again, in `, the
algorithm involves finding the (i + 1)th plane from the remaining 3D range data.

j Ax þ By þ Cz þ D j \TH2 ð 11 : 9 Þ

A threshold value TH1 is affected by the number of planes detected. The lower the
value of TH1, the more iterations of plane detection there are. Therefore, TH1 is
determined by the number of sampling data points and the voting value that differ-
entiates the plane and nonplane cases via the experiments. And, the threshold value
TH2 is the residual of all the points and the plane we have detected. It is also
experimentally determined by the rates of correct data and noises via the experiments.

11.4 Experiments

11.4.1 Range Sensor

We have implemented the proposed method on a Core TM 2 CPU 2.0 GHz computer
and used a Bumblebee stereo imaging system [ 12 ] for obtaining the 3D range data.
The quality of the range data from the bumblebee stereo camera is lower than
that from the active sensor such as laser scanner and it cannot obtain the range data
in a nontexture region, because it fails to find the correlations between the left and
right images. However, the range data from the stereo camera can be obtained
from a low-cost computer practically in real-time.
Figure 11.8 a, b shows the left and right images of the bumblebee stereo camera
and Fig. 11.8 c shows the depth information from the horizontal disparity between
the two images. The region close to the camera is dark, and the bright region is
distant from the camera.
In the right-top region of Fig. 11.8 c, the gray color region doesn't have any
disparity information because there is no texture.

11.4.2 Comparison Results

In the conventional RHT method the sampling times are 2,000,000 per iteration
and planes are detected with the TH1 value of 30 for the IRHT algorithm in
Fig. 11.7 .