Knowledge extraction from visual pattern recognition systems with faster adaptation to new and changed scenarios as goal
The overall goal is to guide the engineer by providing algorithmic approaches of similar inspection problems, and to unburden the engineer from recurrent design, tuning and implementation steps
This requires a better automatic reuse of knowledge and software modules for accelerating the whole engineering process of modelling and algorithmic workflow design.
The objectives address the usability of such an engineering software framework, its productivity and the performance of the resulting software solution in terms of classification rate, robustness, need of processing power and configuration effort.
Aims / Research Topics
Research topics in this project are derived from applications such as traffic monitoring, observation of outdoor sports and similar scenarios. These scenarios are characterized by a high variability, caused e. g. by different weather conditions and the resulting different lighting conditions, variable background traits, etc.
Goals include, among others, the
- reduction of training effort by creating and calibrating virtual reference data from similar scenarios,
- improved robustness of pattern recognition systems by using an improved training database,
- development of an approach to create a hierarchy of higher-ranking concepts for image properties, as well as
- design and optimization of different learning approaches (machine, deep learning).
Methods / Software / Proof of Concepts
- Development of methods for better object recognition, e. g. in the field of traffic or sports.
- Prototype for tracking players and balls in recordings of soccer training and soccer games.
- Prototype for tracking of nanoparticles and subsequent size estimation in videos of a laser scattered light microscope in the range of 20nm and larger.