SIMP3: Social Interaction-Based Multi-Pedestrian Path Prediction By Self-Driving Cars

Abstract

An accurate and fast prediction of future positions of pedestrians by a self-driving car in critical traffic scenarios remains a challenge. The intention of a pedestrian to cross the street can be influenced by social interactions with another one across the street, which may be manifested through various types of social signals such as hand waving. Current socially-aware multi-pedestrian path predictors mainly rely on geometric heuristics such as the distance between pedestrians in the field of view of the car, but do not consider their social interaction across the street. This paper presents a novel social interaction-based multi-pedestrian path predictor (SIMP3) which leverages a combination of dynamic Bayesian networks for intention detection and recurrent network for prediction of future pedestrian locations. The system has been evaluated on the benchmark OpenDS-CTS2 of critical traffic scenarios with socially interacting pedestrians across the street simulated in OpenDS. Our experiments revealed that in most scenarios SIMP3 can significantly outperform the selected competitors.