|
A region-based approach for efficient detection of
unstructured road boundaries is proposed. The proposed approach
applies pattern recognition techniques to detect the boundaries
on a variety of unstructured roads. This is an unsupervised
classification method that detects difficult roads without
tracking results from image to image. In each new image, it
quickly clusters pixels having similar colors into regions using
a combination of ISODATA algorithm and Minimum distance
classification. At the end of the clustering, each pixel in the
image is labels a class, forming a class image. The edges
between class labels in the class image are then extracted using
Kirsch operator, and collected in a class edge image. Finally,
extract the road boundary locations from the class edges using
the straight-line or curve fitting. With a lot of sequential
road-test color images on campus and rural roads under varying
illumination conditions (including strong shadow and noise
conditions), the processing time is short enough to drive the
vehicle for road following in real time. Our approach does not
rely on the assumption of road candidate interpretations (such
as SCARF or UNSCARF), and it detects the road boundaries well.
Comparing with other road-detection systems based on color
image, the proposed scheme is a great improvement in the aspect
of processing speed as well as the reliability of locating the
road boundary.
|