Image Understanding and Scene Analysis

The term "video and scene analysis" refers to the semantic analysis of image data, often recorded outdoors. A simple example is traffic sign detection. Often, however, they are recordings from different sources.

One of the largest projects of the image processing department is the automatic detection of objects photographed with different mobile phones. For the purpose of the largely parameter-free recognition of different objects, current research algorithms are adapted as well as new algorithms developed.

For many years machine learning has been an integral part of many projects and research activities in the department. In the field of surface inspection, hybrids from the "classical" parameterizable methods (filters, morphology, edge detectors) and learning approaches are increasingly being used.

Learning methods, such as "deep learning", require a large number of annotated data, which in an industrial project is usually neither affordable nor practical. For this reason, it is necessary to model assumptions about the objects, you want to locate, and to use this modeling as a partial input for automated procedures. Because of this special data situation, the department's learning approaches are almost always model-based (model-based machine learning).

Overview Example Projects


Action Recognition

Actions of people are an important part of feature films and videos. The automatic recognition and assignment of these actions is an essential component of content-based video analysis systems.



The aim of the project is the automatic detection of the position of relevant objects in the road area, e.g. traffic signs, traffic lights, road markings, and protection barriers.


Video Detection and Retrieval

Together with project partners from the BR and the company AVID, our department has developed an automatic video decoder, which allows you to find the scenes of individual video clips in TV broadcasts.


Analysis of Hyperspectral Images