Physics Informed Deep Learning for Motion Prediction in Autonomous Driving

Abstract:
This paper examines a physics-informed artificial neural network and its ability to learn to forecast car motion sequences using recorded real-world traffic sequences from the Argoverse dataset. It implements the Intelligent Driving Model (IDM), which models straight driving and is derived analytically and combines a data-driven model, a Long-Short-Term Memory (LSTM) neural network at the loss function level. Three networks, the standalone IDM, the LSTM Network, and their combination, a Physics-Informed Neural Network (PINN), are examined and compared in their performance and convergence rate, as well as inspected at the level of individual, visualized traffic scenes. Furthermore, this work covers the implementation details and challenges for the IDM, the combined Physics-Informed Neural Network, and how the data was prepared. We find that the inclusion of the simpler IDM into the training of the LSTM network yields a better initial performance as well as a stabilized model performance during training, we also find, however, that the IDM requires more extensive integration and preparation for working with the chosen Dataset since its simplicity does not lend itself to easily provide applicable collocation points and falls behind a purely data-driven approach.