Probabilistic, Variable and Interaction-aware Situation Recognition

• Main Author: Klingelschmitt, Stefan
• Format: Others
• Language: en
• Published: 2018
• Online Access: http://tuprints.ulb.tu-darmstadt.de/8066/7/2018-09-10_Klingelschmitt_Stefan.pdf; Klingelschmitt, Stefan <http://tuprints.ulb.tu-darmstadt.de/view/person/Klingelschmitt=3AStefan=3A=3A.html> : Probabilistic, Variable and Interaction-aware Situation Recognition. Technische Universität, Darmstadt [Ph.D. Thesis], (2018)

Future advanced driver assistance systems (ADAS) as well as autonomous driving functions will extend their applicability to more complex highway scenarios and inner-city traffic. For these systems it is a prerequisite to know how an encountered traffic scene is most likely going to evolve. Situation recognition aims to predict the high level behavior patterns traffic participants pursue. Thus, it provides valuable information that helps to predict the next few seconds of a traffic scene. The extension of ADAS and autonomous driving functions to more complex scenarios poses a problem to state-of-the-art situation recognition systems due to the variability of the encountered scene layouts, the presence of multiple interacting traffic participants and the concomitant large number of possible situation classes. This thesis proposes and discusses approaches that tackle these challenges. A novel discriminative maneuver estimation framework provides the possibility to assess traffic scenes with varying layout. It is based on reusable, partial classifiers that are combined online using a technique called pairwise probability coupling. The real-world evaluations indicate that the assembled probabilistic maneuver estimation is able to provide superior classification results. A novel interaction-aware situation recognition framework constructs a probabilistic situation assessment over multiple traffic participants without relying on independence assumptions. It allows to assess each traffic participant individually by using maneuver estimation systems that determine complete conditional distributions. A real-world evaluation outlines its applicability and shows its benefits. The challenges associated with the increasing number of possible situation classes are addressed in two ways. Both frameworks allow to reuse classifiers in different contexts. This reduces the number of models required to cope with a large variety of traffic scenes. Moreover, a situation hypotheses selection scheme provides an efficient way for reducing the number of situation hypotheses. This lowers the computational demands and eases the load on subsequent systems.