FRVO: A Filter Enhanced Interaction Model for Pedestrian Path Prediction in Crowded Scenarios

The paper presents a filter enhanced interaction model to predict pedestrian trajectory in crowded scenarios, which is essential for robots navigating in pedestrian-rich environments. An adaptive interaction model based on reciprocal velocity obstacle is employed to simulate interaction among pedestrians. It allows an agent to be aware of other agents' aggressiveness and environment crowdedness. Besides, we implement a filter-based online learning framework with adaptive noise covariance to continuously refine the interaction model's inner state, which addresses the uncertainty of pedestrian path planning. We highlight the path optimality and energy efficiency of our method by simulation. Experiments with real-world dataset demonstrate the generality to unknown environments of our method and at least 25% improvement on path prediction error compared with prior models.