TY - CHAP A1 - Dietl, Laura A1 - Facchi, Christian ED - Zöldy, Máté T1 - Is Maximum Entropy Deep Inverse Reinforcement Learning Suitable for Pedestrian Path Prediction? T2 - Proceedings of the 3rd Cognitive Mobility Conference UR - https://doi.org/10.1007/978-3-031-81799-1_20 Y1 - 2025 UR - https://doi.org/10.1007/978-3-031-81799-1_20 SN - 978-3-031-81799-1 SP - 215 EP - 229 PB - Springer CY - Cham ER - TY - THES A1 - Dietl, Laura T1 - Inverse soft actor-critic imitation learning for pedestrian trajectory prediction N2 - Pedestrian trajectory prediction has the potential to enhance safety in everyday traffic situations. By utilizing Vehicle-to-Everything (V2X) communication pedestrians will be able to communicate their intended future trajectory with other road users. Thus, enabling other road users to adapt their path accordingly, e.g., to avoid collisions. Due to the inherent multimodality of future paths, the impact of the social and environmental context, and the influence of personal factors on a person’s behavior, pedestrian trajectory prediction is a difficult endeavor. In light of these challenges, this thesis’ objective was to develop a pedestrian trajectory prediction model capable of generating realistic multimodal trajectories for environments that closely resemble everyday traffic scenarios. This was achieved by training a model using the Inverse Soft-Q Learning (IQLearn) algorithm, which adapts the Soft Actor-Critic (SAC) approach for use in addressing Inverse Reinforcement Learning (IRL) problems. As far as the author is aware, this was the first time of IQLearn being applied for path prediction. For model training the Stanford Drone Dataset (SDD), a prominent real world dataset, was employed. The dataset analysis reveals characteristics that must be considered during the development and evaluation of a model. Furthermore, a novel difficulty classification scheme helps to gain insight into the pedestrian paths’ complexity distribution within the dataset. It is based on the trajectories’ velocity, heading, number of starts and stops, and the newly developed social context measure, which gauges the level of attention a person must pay to their local neighbors. It shows that the SDD mostly contains simpler trajectories with only a small portion being complex and containing outlier behavior. Evaluating the model on the SDD shows that it: can compete with popular benchmark models based on common metrics; is able to predominantly predict three general types of paths, i.e., small bends, linear and stationary behavior, indicating that it learned a generalized representation of the trajectories within the training dataset; and is not able to generate multimodal paths, as it predicts future samples with only minor deviations from one another instead of a diverse range of realistic trajectories. The dataset, which contains primarily simpler paths, lacks the required complexity for the model to anticipate outlier behaviors. Instead, it often completes the trajectories in a logical manner, adhering to the constraints of the environment. Y1 - 2024 UR - http://nbn-resolving.de/urn/resolver.pl?urn:nbn:de:bvb:573-53401 CY - Ingolstadt ER - TY - CHAP A1 - Dietl, Laura A1 - Facchi, Christian T1 - Really, Pedestrian Trajectories: How Realistic are the Datasets? T2 - 2025 IEEE Intelligent Vehicles Symposium (IV) UR - https://doi.org/10.1109/IV64158.2025.11097441 Y1 - 2025 UR - https://doi.org/10.1109/IV64158.2025.11097441 SN - 979-8-3315-3803-3 SP - 301 EP - 308 PB - IEEE CY - Piscataway ER - TY - CHAP A1 - Dietl, Laura A1 - Facchi, Christian T1 - A Glimpse into the Future: An Inverse Soft Q-Learning's Soft Actor-Critic Approach for Pedestrian Path Prediction T2 - 2025 IEEE Intelligent Vehicles Symposium (IV) UR - https://doi.org/10.1109/IV64158.2025.11097431 Y1 - 2025 UR - https://doi.org/10.1109/IV64158.2025.11097431 SN - 979-8-3315-3803-3 SP - 111 EP - 118 PB - IEEE CY - Piscataway ER -