Interpretable and Realtime Predictions of Social Interactions for Autonomous Vehicles

Abstract

Autonomous vehicles present an opportunity to transform transportation. The benefits range from increased access to mobility and time freed from driving, to greater safety due to automation. These robots are powered by the coordination of various systems to perceive and navigate through the world. Crucially, the autonomous vehicle operates in an open environment alongside fellow road users with whom it will interact regularly. Predictions of fellow road users' intents and future motion guide these interactions and specify a large part of the autonomous vehicle's behavior. Spurred by advances in deep learning, recent prediction methods have increasingly begun to consider how interactions affect future motion in ever more varied environments. The corresponding gains in accuracy translate to more anticipatory and less reactive autonomous vehicle behavior. One drawback is an increase in complexity, which can lead to less interpretable predictions and behavior. Achieving realtime performance and handling missing data caused by adverse sensing conditions present additional challenges.

To support autonomous vehicle behavior that is transparent and predictable, this thesis develops interpretable prediction methods. Model-based approaches provide the vehicle for building interpretable predictions, and novel inference procedures are developed to generate the predictions in realtime. Adopting a probabilistic framework enables natural handling of missing data and affords the flexibility to model interactions in varied environments beyond those described by existing interpretable methods. Experiments on real highway traffic and urban data demonstrate the developed methods' effectiveness.