- Abstract: In this work we build a model for pedestrian encounters parametrized by aggressiveness and attentiveness based on the SHRP2 dataset, which includes dashcam and driver facing video footage from several different regions within the US. From this dataset we use inverse reinforcement learning to extract a reward function to model pedestrian actions parametrized by attentiveness/distractedness and passivity/aggressiveness. The dataset is parsed and labeled by a video analytics toolkit we develop. Finally, we use these models to design an autonomous driver that makes optimal decisions according to a tunable parameter of desired aggressiveness.
- TL;DR: Use database of dashcam videos to learn pedestrian behaviors when in unmarked crossing scenarios.
- Keywords: Autonomous Vehicles, Inverse Reinforcement Learning, Modeling