· Contributors · Organizations ·
Exploring Discrepancies between Drivers and AI in Understanding Pedestrian Intent
SessionPoster Session 1
DescriptionWith the fast progress of artificial intelligence, autonomous vehicles (AVs) are expected to get into the daily lives of normal people. However, the advanced learning algorithms used in AVs cannot predict pedestrian behaviors accurately, creating challenges to implementing fully autonomous AVs in urban settings. To better understand the social intelligence level of modern AVs, this study compares human drivers and AI algorithms with the common task to estimate Pedestrian Situated Intents. Through conducting a machine behavior study with video experiments, the results show that state-of-the-art pedestrian behavior prediction algorithms behave quite differently from human drivers, with similar pedestrian intent estimation results found in <4% of all 110 randomly sampled encountering scenes. Different patterns of human-AI discrepancies highlight the limitations for AVs to perform social actions and predict pedestrian behaviors.