B1387
Title: Bayesian methods for informing trajectory predictions in safe autonomous driving
Authors: Christian Schlauch - Humboldt Universitaet zu Berlin (Germany) [presenting]
Christian Wirth - Continental AG (Germany)
Nadja Klein - Karlsruhe Institute of Technology (Germany)
Abstract: Ensuring the safety and trustworthiness of autonomous driving systems demands probabilistic, multi-modal trajectory predictions that can reliably handle unexpected and complex scenarios. However, current deep-learning prediction models tend to be brittle and often fail to consider crucial information, such as road topology. In a Bayesian framework, expert knowledge can be used to make this information explicit and inform the prediction models. Unlike existing informed learning approaches, the probabilistic informed learning approach does not require any model architecture changes or specific knowledge representations. Applied to two state-of-the-art prediction models, a substantial increase in accuracy and robustness is demonstrated, as evaluated on the public NuScenes dataset. These findings highlight the potential to enhance safety-critical applications where valuable expert knowledge is readily available.
