Datenbestand vom 09. März 2024

Warenkorb Datenschutzhinweis Dissertationsdruck Dissertationsverlag Institutsreihen     Preisrechner

aktualisiert am 09. März 2024

ISBN 9783843940009

72,00 € inkl. MwSt, zzgl. Versand


978-3-8439-4000-9, Reihe Informatik

David Lenz
Motion Planning for Highly-Automated Vehicles under Uncertainties and Interactions with Human Drivers

141 Seiten, Dissertation Technische Universität München (2018), Softcover, A5

Zusammenfassung / Abstract

This thesis addresses the problem of motion planning for automated vehicles considering interactions with human drivers and uncertainties from perception and actuation based on the example of lane merging scenarios in dense traffic. In these scenarios, it is not sufficient to first predict all other traffic participants without regarding the motion of the ego-vehicle and then plan based on this prediction, because the behavior will be too passive. When explicitly considering the reaction of other cars during planning, two problems arise: First, a model for the reaction - which is non-deterministic - is necessary that can be calculated efficiently to allow real-time planning. Second, the state space of the problem increases drastically which prohibits many classical planning techniques.

These two problems are addressed by employing the anytime algorithm Monte-Carlo Tree Search and by training interactive models based on Neural Networks. The models describe the reaction of other traffic participants and are additionally utilized as a policy to guide the MCTS search process to focus the computational power to interesting regions of the state space. With this technique, it is possible to plan these complex problems in real-time.

Based on real scenarios taken from the NGSIM dataset it is shown that the proposed planner is able to generate reasonable human-like behavior during lane merging considering the cooperation and interaction with other vehicles, increasing the lane change success rate compared to a baseline planner. Furthermore, the convergence of MCTS is shown to be improved by the guidance of the trained model.

A final stochastic nonlinear optimization makes the high-level trajectory drivable, comfortable, and safe considering the capabilities of an underlying control algorithm and uncertainty of the prediction. This is achieved by formulating the problem as a Gaussian system with a cost term based on risk, that minimizes the probability and severity of a collision.