Detailseite
Projekt Druckansicht

Active Random Hypersurface Models: Simultane Form- und Lageschätzung ausgedehnter Objekte in verrauschten Punktwolken

Fachliche Zuordnung Automatisierungstechnik, Mechatronik, Regelungssysteme, Intelligente Technische Systeme, Robotik
Förderung Förderung von 2013 bis 2016
Projektkennung Deutsche Forschungsgemeinschaft (DFG) - Projektnummer 234520279
 
Erstellungsjahr 2017

Zusammenfassung der Projektergebnisse

In this project, we proposed and developed an approach to estimate extended target objects based on point measurements. The result was called Active Random Hypersurface Models, which extends the modeling techniques of Random Hypersurface Models with ideas drawn from Active Contour Models. The key idea was to combine state-of-the-art association models based on probabilistic association and distance-based minimization, in order to construct complex target shapes from simple forms using extrusions, symmetries, compositions, and other transformations. In this context, we planned to solve three challenges. First, the resulting estimator needed to handle different uncertainties specific to each measurement, which could change quickly in time. Second, the shape models needed to be flexible, allowing for low details in cases of low measurement quality, and high accuracy when more information is available. Finally, it should be able to deal gracefully with low prior information. In particular, it should allow for the estimation of shape and pose simultaneously. The project had to undergo significant restructuring due to the fact that its work packages were planned for two researchers, but the final grant provided funds for only one researcher. Because of this, we planned five blocks that encompassed about half of the tasks of the proposed work packages. The first block was concerned with the sensor model. Due to the modified time constraints, a new sensor framework could not be developed, and the models used were based heavily on previous work. Instead, we focused on exploiting all the available sensor information by incorporating negative measurements, which told us where the target cannot be, in addition to the traditional positive observations. For the second block, we dealt with modeling simple, two-dimensional shapes. To achieve this, we proposed a bias-reduction approach that allows for shape estimation even in high noise environments. The third block was concerned with describing arbitrary, non-convex shapes, an issue which was addressed with the introduction of Level-set ARHMs. For this approach, the idea was to describe the interior of a shape through level-sets of a distance function, and to represent the boundary by polygons subject to regularization. The fourth block developed a technique to translate the previously explored concepts into three dimensions, with the use of extrusions and symmetries. Finally, a mechanism for composition, that is, describing a shape as the union of multiple components joined together, was explored in the fifth block. Overall, the main result of the project was a recursive Bayesian estimator for tracking a wide variety of extended objects, which can be used in both particle filters and nonlinear Kalman filters, and easily allows for the fusion of multiple types of measurements. This estimator can find applications in a broad spectrum of fields, such as robotics, indoor navigation, and augmented reality. We also showed two possible direct applications on projects on which the ISAS laboratory is participating. While not all work packages could be implemented, they can still benefit directly from the work presented by this project, and can be implemented as part of a future grant.

Projektbezogene Publikationen (Auswahl)

 
 

Zusatzinformationen

Textvergrößerung und Kontrastanpassung