PhD defence of Jilan Samiuddin – Decision making in self-driving cars using Graph Neural Networks

Abstract

The autonomous driving industry's rapid growth highlights the necessity for advanced technologies to guarantee safety, comfort, and efficiency. This thesis focuses on three fundamental aspects of autonomous driving systems: trajectory prediction, trajectory planning, and control adaptation. The first contribution of this study is the introduction of a new technique for trajectory prediction that utilizes spatial-temporal graphs to capture historical traffic interactions. The use of a depthwise graph encoder network and sequential Gated Recurrent Unit decoder improves vehicle trajectory prediction compared to other deep learning methods. Next, an innovative online graph planner is introduced for generating feasible and comfortable trajectories. The planner creates a spatial-temporal graph that integrates the autonomous vehicle, nearby vehicles, and virtual road nodes. The graph is then processed using a sequential network with a behavioral layer for kinematic constraint compliance. Testing the planner on complex driving tasks demonstrates its effectiveness, surpassing existing state-of-the-art approaches. Finally, a novel approach for online learning in vehicle modeling and lateral control is introduced, using heterogeneous graphs and Graph Neural Networks. This technique enables the vehicle model and lateral controller to adapt to dynamic conditions, enhancing performance under perturbations. The self-learning model-based lateral controller is evaluated on the CARLA simulator, showing promising results. These contributions improve trajectory prediction, planning, and control adaptability, advancing autonomous driving technology and enhancing safety and efficiency of autonomous vehicles.