FO-ORCA: An Optimization-Enhanced Interaction Model with Filtering for Pedestrian Trajectory Prediction in Dense Crowds

The coexistence of humans and robots in dynamic environments relies on the ability of intelligent agents to understand each other's behaviors and predict their movements. This paper proposes an optimization-enhanced interaction model for pedestrian trajectory prediction in crowded environments, which is crucial for robot navigation in densely populated settings. The model, based on Optimal Reciprocal Collision Avoidance (ORCA), simulates interactions among pedestrians. It enables an agent to recognize the responsibility for making minor changes in both the velocity direction and magnitude when interacting with other agents, reflecting the characteristic of pedestrians seeking the most efficient path to their destination. Furthermore, we introduce an intention-driven online learning framework that incorporates probabilistic decisionmaking, model parameter learning, and noise variance adaptation. This framework continuously refines both the internal and external interaction states of the model, addressing the uncertainties in pedestrian path planning. The simulation results validate the optimality and energy efficiency of the generated paths. Experiments on real-world datasets demonstrate the generalizability of our approach to realistic environments, with improvement in trajectory prediction accuracy compared to existing models.