PRECISE DETECTION AND REPRESFNTATION OF LINEAR FEATURES
Jan Heikkilä
Helsinki University of Technology
Institute of Photogrammetry and Remote Sensing
Finland
Abstract
A system for detecting linear features, like roads and object boun-
daries, with high precision, is presented. The algorithm combines
three well-known operators. First, the subpixel detection of can-
didate edge pixels, is done by the Nalwa-Binford operator. Second,
the corner points are detected by the Fôrstner operator, but loca-
ted as the intersections of the Nalwa-Binford edges. The final
representation is done with parametric B-splines, which are fitted
under the L,-norm to the data. This description can cope with
local rank-defiencies and discontinuties in the data.
1.0 INTRODUCTION
Detection of linear features plays a central role in the automation
of photogrammetric procedures. This is true both in conventional
(cartographic) applications and in the new applications of close-
range photogrammetry (machine vision). Interesting future photog-
rammetric systems could be based purely on linear features (c.f.
/MulMik88/) offering more reliable and flexible measurement proce-
dures.
In the photogrammetric applications the measurements should usually
be done with high precision. Concerning the use of state of the
art digital imagery this means subpixel accuracies. In the photog-
rammetric community much attention has been paid on the high pre-
cision measurements with the help of pointlike features. However,
linear features actually play a more important role in the automa-
tion process. Many physical processes can be represented by linear
features, e.g. discontinuities in object geometry and illumination
(image and scene segmentation problems). Area based methods (c.f.
/Nevati86/) are an alternative for the segmentation problem. Howe-
ver, they cannot produce subpixel accuracies. That is why linear
feature detection is needed, at least as a post-processing stage.
In the combination of the global methods for surface recovery (c.f.
/Wrobel87/, /EbnHei88/, /Helava88/, /Li89/) linear features also
offer a possibility to model (detect and represent) the surface
discontinuities precisely.
286
N
b E di NS rnm
a3
ts: 132 IM ALM ud
