Combining Stereo Vision with Topological Constraints to Solve Knowledge based Object Recognition Tasks

E. Büscher (Germany)

Keywords

Distributed Artificial Intelligence –Knowledge Acquisition - Image understanding - Neural Networks

Abstract

Since3-D and cartographic data becomes more and more important for several applications based on or producing accurate 3D models, this contribution focuses on the knowledge based automatic extraction and modeling of buildings from overlapping views of urban scenes. In order to reach that goal we apply a method that uses both statistical and structural information. In addition several heuristic rules are added in order to improve the effectiveness of these methods. The rules are based on gradient correlation, contour-adaptive windows and a two-way filtering so that accurate, reliable, and discontinuity-preserving digital elevation models (DEM) are one result of the whole expert system. Based on the parameterized model approach of recent works, the most difficult task of establishing appropriate correspondences between model and image features becomes feasible. Thresholds are involved with respect to the physical characteristics of the scene where inductive learning is applied. The inductive algorithm trains on a subset of the DEM that has been hand-classified, then automatically classifies the primitives in the remaining (unclassified) portion of the scene. Artificial feed-forward neural networks (ANN) are a logical choice for consideration of this task of learning features in digital images and DEMs. Experiments with several objects show the usability of the proposed approach.

Important Links:



Go Back


IASTED
Rotating Call For Paper Image