Andrew Bibitchev
e^, e?) Gn) 6?) 0^) Qn) am
Therefore on layer L—1 of the pyramid maximization (14) can be performed only in vicinities of 2d edges
(a? ); bs ) and (a! à 6? à, , that allow us to decrease area of search radically. As a result we obtain set of 3d
edges {at i ph! } M . And so on until bottom level of the pyramid will be reached.
Due to Gauss pyramid application the speed of the algorithm goes up in times.
4 APPLICATION EXAMPLES
Initially, proposed technique was developed for building extraction from aerial imagery. For this reason application
examples are auto and semi-auto buildings extraction.
Available input data is:
— stereo-pair of gray scale aerial images;
— parameters of internal, relative and external orientations;
- some parameters of buildings (minimum and maximum height, minimum and maximum length of short side and so
on);
— DEM produced by stereo-correlation procedure (optionally).
Required output data is:
— 3d digital models of building roofs;
— DEM of ground (optionally).
4. Semi-Auto Building Extraction
The main idea of semi-auto approach consists in the following. User performs only supervisor functions, and the
computer does all stale and accurate operations. The approach scheme is presented in Figure 7.
| Len USER iso j 3d digital
set of models of roofs
building ^ Areas of "
1. Selection of oor 2.User sets key points 3.Calculation of | interest |4.3d edges edges 5.3d edges
building type points on one of the » interest areas » extraction » grouping and
(from list) available images on each image selection
i
=>: 2d model of roof image
3d model of roof
Figure 7. Scheme of semi-auto building extraction procedure
First of all user must select type of the building. For each building type appropriate model is supposed to be known.
Then, user must mark on any image rough position of key points. Key point selection is some kind of “know how” and
depends on building type. For example, in case of rectangle flat roof we can use two key points: one arbitrary point on
each short side of the roof image. Next, using marked positions of key points we can produce areas of interest on each
available image. In these areas algorithm of 3d edge extraction is run. As result we obtain candidates to roof sides.
Using key points and 3d relational model of the roof selection and grouping of 3d edges are performed. Thus, we obtain
a number of roof hypothesizes. The hypothesis with maximum total weight is selected as a final result of reconstruction.
Note that total weight of roof hypothesis must include as weights of 3d edges constituting roof as measure of adequacy
to the model.
Described approach is used commercial Win32 application “Simple Building Extraction”, which allows easy and fast
create accurate 3d digital models of cities and towns.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 77