Autonomous vehicles may be driven by image data of real-world scenes c
ollected through a TV camera. Detecting the clues for safe navigation
requires, among other things, the estimation of the path to be followe
d by the vehicle, which has proven to be a formidable task in outdoor
scenes. In this paper, an innovative system for road direction detecti
on is proposed which is composed of three specialized blocks performin
g edge extraction, image-segments detection and road estimation. The r
oad direction estimation block is implemented as a feedback neural net
work and is not fed directly with image data but with higher-level ima
ge features which are extracted through the preprocessing stages. The
use of feedback, while reducing the complexity of the network, improve
s the estimation robustness and the noise immunity. A novel algorithm
is defined and employed for the training step and experimental results
in outdoor scenes are reported.