Third party funded individual grant
Acronym: EAAFLS
Start date : 15.05.2016
End date : 31.03.2017
Website: https://www2.cs.fau.de/research/EAAFLS/
The goal of this project is the development of adaptive algorithms for radio-based realtime locating systems. In the scope of this project we cover three essential topics:
Automatic configuration of event detectors. In previous research projects we built the basics for the analysis of noisy sensor data streams. However, event
detectors still need to be parameterized carefully to yield satisfying results. This work package explores the possibilities of an automatic configuration of the event detectors on existing sensor and event data streams.
In 2016 we investigated concepts to extract optimal configurations from available sensor data streams. For soccer sport scientists manually annotated matches and scenes (e.g. player A kicks the ball with his/her left foot at time t). These manually annotated scenes may later by used to optimize the hierarchy of event detectors.
Evaluation of machine learning techniques for locating applications. In previous research projects we already developed machine learning algorithms for radio-based locating systems (e.g., evolutionary algorithms to estimate antenna positions and orientations). This work package investigates further approaches that use machine learning to enhance the performance of realtime locating systems.
In 2016 we evaluated concepts to replace parts of the position estimation algorithms by machine learning algorithms. Up to now a signal processing chain (analog/digital conversion, time of arrival estimation, Kalman filtering, motion estimation) uses the raw sensor data to calculate a position. This often results in high installation and configuration costs for the setup of locating systems in the application environment.
Evaluation of vision-based techniques to support radio-based realtime locating.
Radio-based locating systems have strengths if objects are occluded as microwaves may pass through the occluding objects. However, metallic surfaces in the environment pose challenges as they reflect RF-signals. Hence, the RF-signal that a transmitter emits arrives at the antennas over multiple paths. It is then often difficult to extract the directly received parts of the signal at the antenna and hence it is a challenge to properly estimate the distance between the antenna and the emitter. In this work package we investigate vision-based locating techniques that may help RF-based systems in calculating positions.
In 2016 we developed two systems: CNNLok may be used by objects carrying a camera (self-localization), i.e., inside-out tracking, whereas InfraLok uses cameras installed in the environment to track objects with infrared light. CNNLok uses a convolutional neural network (CNN) that is trained on several camera images taken in the environment (at known places). At runtime the CNN receives a camera image and calculates the position of the camera. InfraLok detects infrared LEDs using a multi-camera system and calculated the position of objects in space.