NIAR: Interaction-aware Maneuver Prediction using Graph Neural Networks and Recurrent Neural Networks for Autonomous Driving

Human driving involves an inherent discrete layer in decision-making corresponding to specific maneuvers such as overtaking, lane changing, lane keeping, etc. This is sensible to inherit at a higher layer of a hierarchical assembly in machine driving too, in order to enable tractable solutions for the otherwise highly complex problem of autonomous driving. This has been the motivation for this work that focuses on maneuver prediction for the ego-vehicle. Being inherently feedback structured, especially in dense traffic scenarios, maneuver prediction requires modeling approaches that account for the interaction awareness of the involved traffic agents. As a direct consequence, the problem of maneuver prediction is aggravated by the uncertainty in control policies of individual agents. The present paper tackles this difficulty by introducing three deep learning architectures for interaction-aware tactical maneuver prediction of the ego-vehicle, based on motion dynamics of surrounding traffic agents. Thus, the traffic scenario is modeled as an interaction graph, exploiting spatial features between traffic agents via Graph Neural Networks (GNNs). Dynamic motion patterns of traffic agents are extracted via Recurrent Neural Networks (RNNs). These architectures have been trained and evaluated using the BLVD dataset. To increase the model robustness and improve the learning process, the dataset is extended by making use of data augmentation, data oversampling, and data undersampling techniques. Finally, we successfully validate the proposed learning architectures and compare the trained models for maneuver prediction of the ego-vehicle obtained thereof in diverse driving scenarios with various numbers of surrounding traffic agents.