Door Identification -- Object Extraction
Given the B/W image produced by the filtering stage, we segment the
white pixels into objects using a labeling algorithm. For this task we
use the MATLAB function bwlabel that labels pixels by running a
window of a given size over the image, labeling two points with the
same label if they are in the same window. After labeling the pixels,
we separate the largest objects using the MATLAB function
imfeature (with the 'Area' function) which computes the "mass"
of the object in terms of the number of pixels in it. (For efficiency,
we then crop each object so that the images used in the rest of the
algorithm are significantly smaller.) The output of the algorithm
after processing the image from Figure 2 (those objects which
are fed to the rest of the algorithm) is shown in the following
figure:
 |
|
 |
| Figure 5: Superimposed objects with mass >= 1000 (0.5*size) |
|
Figure 6: Superimposed objects with mass >= 250 (0.5*size) |
Although not significant in these figures, there are several problems
that this process may encounter or introduce. The more common problem
is that different objects may be "glued" together. This especially
happens with noise from the floor which can get associated with the
door.
A more serious problem may occur if for some reason an object
gets split into several smaller objects. In this case our treatment is
risky, because further computation will not associate features of the
different parts together (see experiment
203b). Nevertheless, the assumption that the door
is a continuous surface that is not obscured to such a dangerous
extent seems plausible and postprocessing may correct "broken" doors.
Last, one may want to adjust the "mass"
threshold to different values for different environments. It seems
that for our environment the values of 250-1000 are good enough and clear
some of the noise away yet keeps even distant doors. The experiments
we show at the experiments section we
produced using the mass threshold of 250.
Next: Shape finding
Return to Table of Contents
Eyal Amir
and Pedrito Maynard-Reid II.
Last modified: Tue Mar 16 19:04:41 PST 1999