An Approach to Knowledge-Based Interpretation of Outdoor Natural Color Road Scenes

This paper describes an approach to robust road scene interpretation in high-level vision. The two key ideas are adjustable explicit scene models, and an interpretation cycle of evaluation, modeling, and extrapolation. The interpreter first picks up some tractable segment regions and generates initial hypotheses, then iterates its interpretation cycle until every region is labeled. In each interpretation cycle, labeled regions play a survival game. Those which are consistent get their plausibility value increased. Those which are not get their plausibility value decreased and some of them die (become unlabeled). Surviving regions propagate the interpretation to neighboring similar regions. In the mean time, explicit scene models are also adjusted with the latest interpretation in each cycle. The models gradually extract global information from the scene, for instance hypothesizing the location of the road. When the plausibility values of the models become high, the system begins global evaluation, using the models, in addition to the local evaluation. that uses region features and adjacencies. The combination of global and local interpretation allows even deceptive regions to eventually be correctly labeled. More than 20 road scenes, some destructively shaded, are reliably interpreted by the system INSIGHT-III which implements this approach. 1