D.3 Multimodal Szene Analysis

Project Leaders

Jun.-Prof. Dr.-Ing. Marcin Grzegorzek, Prof. Dr.-Ing. Andreas Kolb

Initial Situation

Automatic object recognition and semantic scene analysis are still part of the unsolved problems of pattern recognition. Most of the existing methods for scene analysis utilize a single sensor modality for the picture, generally 2D-RGB, rely on a certain method of feature calculation, as e.g. based on appearance or form, and use a given classification procedure. This achieves satisfactory robustness only at the cost of greatly reduced complexity, as e.g. in the case of industrial applications. In contrast, although generic object recognition methods have been defined application independently, in many cases they achieve unsatisfying classification rates.
.

Aims and Work Program

In this project, data from RGB-, depth-, and multispectral sensors are processed and consulted for scene analysis. On the one hand an approach to an adaptive, multimodal object representation and -recognition ist drawn up and realized, and on the other hand a method for the integration of ontological background knowledge into sensor based scene analysis is developed, whereby a semantically advanced interpretation of scenes can result. The prevailing aim of this project is the research of generic object recognition methods, which can profit from context knowledge, (knowledge from within a certain application domain) and from the adaptivity of data presentation, in order to facilitate scene analysis on the highest possible semantic level.