MALE-A: Stimuli and Cause Prediction for Maneuver Planning via Graph Neural Networks in Autonomous Driving

The driving behavior inherently involves different tasks, one such task corresponding to maneuver planning to navigate safely the road network. Yet, the planned maneuver can not be executed all the time due to the presence of other vehicles, traffic signs or traffic rules. This work proposes a two-layer approach for decision-making, where the first layer involves planning the high-level maneuver, while the second layer predicts the stimuli of maneuver execution. To predict the required stimuli action, the deep learning model introduced in this work utilizes inputs such as the planned maneuver, automotive sensor data, and an interaction graph constructed from detected traffic participants and features of the traffic scene. This model is based on Graph Neural Networks (GNNs) for extracting spatial features from the graph, while also tracking the temporal context of the traffic scene using Recurrent Neural Networks (RNNs). Training and validation of the model are performed using the HDD dataset, with the application of data oversampling and undersampling techniques to stabilize the learning process. Ultimately, successful implementation of the proposed two-layer approach and model for decision-making is demonstrated across various complex traffic scenes.