Improving Motion Prediction in Autonomous Driving With Expert Knowledge - A Bayesian Deep Learning Approach

Autonomous driving is one of the most highly anticipated yet elusive mobility innovations. The field has made significant advances through deep learning, especially in perception and motion prediction. Still, the field faces open challenges, since safety requirements in autonomous driving demand for robust domain adaptations between locations and well-calibrated uncertainty estimates for the numerous high-risk edge cases in urban environments.

To meet those demands, we explore the potential of Bayesian deep learning methods to motion prediction. We demonstrate how a Bayesian approach can be used to regularize commonly employed motion prediction models by utilizing prior expert knowledge. More specifically, we adapt a CoverNet baseline model with a compute-efficient last layer Gaussian Process approximation and integrate prior drivability knowledge. Doing so improves both, robustness and calibration.

Organized by:

LMU Center for Advanced Management Studies

LMU Munich School of Management