Petrit Rama, Naim Bajcinca

• The driving process involves many layers of planning and navigation, in order to enable tractable solutions for the otherwise highly complex problem of autonomous driving. One such layer involves an inherent discrete layer of decision-making corresponding to tactical maneuvers. Inspired by this, the focus of this work is predicting high-level maneuvers for the ego-vehicle. As maneuver prediction is fundamentally feedback-structured, it requires modeling techniques that take into consideration the interaction awareness of the traffic agents involved. This work addresses this challenge by modeling the traffic scenario as an interaction graph and proposing three deep learning architectures for interaction-aware tactical maneuver prediction of the ego-vehicle. These architectures are based on graph neural networks (GNNs) for extracting spatial features among traffic agents and recurrent neural networks (RNNs) for extracting dynamic motion patterns of surrounding agents. These proposed architectures have been trained and evaluated using BLVD dataset. Moreover, this dataset is expanded using data augmentation, data oversampling and data undersampling approaches, to strengthen model's resilience and enhance the learning process. Lastly, we compare proposed learning architectures for ego-vehicle maneuver prediction in various driving circumstances with various numbers of surrounding traffic agents in order to effectively verify the proposed architectures.