A distributed 3D scene recognition system based on a multilevel repres
entation of object models and signals is described. The solution to a
recognition problem is obtained through a set of object-observation co
uples at the different abstraction levels. The various system modules
exchange two kinds of information: 1) top-down messages, which are use
d to communicate to lower modules the predictions made on the basis of
a priori knowledge on the application domain, 2) bottom-up messages,
which are used to communicate to higher modules the evidence supportin
g possible local solutions. A local scheme for the combination of mess
age flows is defined, and messages are interpreted by using a probabil
istic network of estimators of random variables. The proposed model is
suitable for addressing the problem of distributed geometric reasonin
g aimed at 3D road scene recognition by an autonomous vehicle. Recogni
tion results include road detection and obstacle localization, togethe
r with a study of the relative computational load required by differen
t modules of the system. The proposed approach is currently simulated
on a workstation, while an effective implementation on board of an aut
onomous vehicle is under development in the contest of the CEC-EUREKA
Prometheus project.