Social Aware Maneuver Prediction based on NeuralODE for Autonomous Driving

Learning-based decisions using behavior modeling is a significant research topic in the area of autonomous driving. The lane change maneuver emerges as a crucial factor in determining the behavior of a vehicle affected by the surrounding actors and creating social awareness. In this paper, we propose an enhanced neural network architecture based on Neural Ordinary Differential Equations (NODEs) for the purpose of accurately forecasting the maneuvers of surrounding vehicles in the operational design domain of highway driving. To validate the effectiveness of our approach, we conduct a comprehensive comparison with alternative neural network architectures, namely Temporal Convolutional Network (TCN), Residual Networks (ResNets), and Convolutional Neural Network (CNN). For the comparison, we train and evaluate our models using NGSIM US-101 and I-80 datasets, with a specific focus on three distinct maneuvers: Lane Keep (LK), Left Lane Change (LLC), and Right Lane Change (RLC). By analyzing the performance of our NODE-based architecture alongside the alternative architectures, we observe that accurate maneuver predictions regarding to ground truth values of dataset.