Multiple Plane Detection Method from Range Data of Digital Imaging System for Moving Robot Applications - Visual Computing: Scientific Visualization and Imaging Systems - page 193

into another parameter space, with q-h-u, representing the vertical distance from
the origin and the rotation angle on the plane for voting. The corresponding planes
are detected by the peaks in the voting space.
Although the method produces segmentation results for range data with some
noises and complex planes for an indoor environment, it requires many impractical
procedures such as excessive memory for voting, decisions about the optimal
voting size, difficulties in fixing the peak locations in the voting space, and many
additional processes which need a high computer processing speed [ 2 ]. In order to
solve these problems, many improved methods have been proposed such as
Combinatorial HT (CHT) [ 3 ], Randomized HT (RHT) [ 4 , 5 ], Probabilistic HT
(PHT) [ 6 ], and Dynamic Generalized HT (DGHT) [ 7 ].
RHT is proposed to solve the computing time problem of HT. It doesn't vote all
points in an image into Hough space but votes on the geometrical parameters
which are calculated by selecting several points randomly. However, sampling
range data that exists on different planes or includes noisy samples on a nonplane
surface result in many local maxima in the parameter voting distribution of the
Hough space. Also, several distributions are mixed and thus the peaks can be very
similar. In a mixed and overlapped distribution of peaks in Hough parameter
space, it is difficult to segment a plane reliably.
Kang et al. proposed the plane detection cell (PDC) for reducing the effect of
noises and outliers [ 8 ]. PDC is a cell of circle type used to test the planarity. The
normal vector of a triangle is inscribed in a small circular region, the triangle is
rotationally sampled and a series of inscribed triangles having different normal
vectors is generated. The direction vectors of these generated triangles are used to
test the planarity of the small circular region.
This method is effective in a local region detection but it is not adequate for
global plane detection and segmentation and calculating the planarity at all the grid
point is too time-consuming.
We test only the normal directions of two planes, which are determined by three
vertices of a triangle on range data and arbitrary rotation of it. If the angle of two
normal detections is lower than a given threshold it is voted into Hough parameter
space. It takes less time than PDC and also detects the global planes, because it can
provide the global voting through the local evaluation of data. First, we start with a
scan window to vote locally, then, the scan window explores all regions as its size
grows. This method improves the detection performance because the planes are
locally consistent. We obtained 3D range data from a stereo imaging system and
experimented with it in various environments.
The remainder of this chapter is organized as follows: Sect. 11.2 describes the
RHT method for plane detection. Section 11.3 presents details of the proposed
method. Section 11.3.1 describes a method using random sampling from accuracy
evaluation of local planar region, Sect. 11.3.2 shows a search method using a
scan window, and Sect. 11.3.3 provides a Look-Up-Table for fast processing.
Section 11.3.4 explains the application of Iterative RHT to reduce the effect of
noises and multiple planes. The experimental results are shown in Sects. 11.4 and
11.5 gives the conclusions.