TrackOpt

The tracking of individually occurring points or dense point clouds is highly relevant in a variety of physical and biological problems. Such problems pose a variety of challenges for machine learning methods, since learning must be possible on the basis of often relatively small amounts of annotated training data that is often subject to measurement uncertainties. The aim of this project is to create an efficient method framework that enables new tracking problems to be flexibly integrated and reliably solved. Model-driven, potentially discrete optimization problems are integrated with deep neural networks so that the strengths of both approaches can be exploited. In particular, the integration of model-driven knowledge means that the planned framework can work in a very data-efficient manner. At the same time, learned networks are optimized to predict consistent solutions despite uncertainty and measurement inaccuracy.

We create a resilient framework on the basis of three fields of application that enables current tracking problems to be flexibly integrated and reliably solved. Specifically, this involves particle physics, microfluidics and microscopy. In order to achieve high data efficiency and simultaneous resilience to measurement errors, model-driven optimization problems, which make it possible to include conditions for a valid solution in terms of the biological or physical problem, are integrated with deep neural networks.

We develop a reusable and transdisciplinary, modular software toolbox that also enables external parties in many research and application areas to independently solve complex tracking problems with different boundary conditions. This will be made available to the interested public as an open source application.

University of Mannheim, Chair of Machine Learning

• Prof. Margret Keuper
• Steffen Jung

Heinrich Heine University Düsseldorf, Chair of Machine Learning

• Prof. Paul Swoboda

University of Siegen, Experimental Particle and Astroparticle Physics Group

• Prof. Markus Cristinziani

University of Siegen, Chair of Computational Sensorics

• Prof. Ivo Ihrke

Ilmenau University of Technology, Group of Engineering Thermodynamics

• Prof. Christian Cierpka
• Sebastian Sachs