A Causal Model for Physics-Conform Vehicle Trajectories

Recent advances in autonomous driving have confirmed the need to further improve the robustness and explainability of machine learning modules in this domain. Especially, it is crucial to avoid errors in perception since it negatively impacts the downstream performance and results in undesirable driving decisions. Hence, there is an interest to model the physical knowledge about vehicle trajectories, so that the physical validity of the perceived trajectories can be constantly checked and ensured. Common prior approaches of sequence modeling do not allow us to analyze causal effects. To address this shortcoming, we propose a causal model based on the structural causal model (SCM) framework that explicitly incorporates the physical knowledge about vehicle dynamics. Non-physical dependencies in the model are trained from a real-world dataset. We demonstrate that our model can be used to answer queries at the three levels of the causal hierarchy. We highlight the application of our model to checking whether a sequence of observations, e.g., provided by a perception module, complies with physical knowledge about vehicle trajectories. We present three different conformity check methods based on our model and demonstrate very reliable performance in detecting unreasonable observation sequences.