Hybrid GRU-PINN Model for Pedestrian Trajectory Prediction at Unsignalized Intersections

Pedestrian–vehicle interactions at unsignalized intersections present significant challenges for trajectory prediction. Purely physics-based approaches lack the flexibility to capture human decision-making, while purely data-driven models often struggle with generalization and physical plausibility. The proposed Hybrid GRU-PINN Framework models pedestrian trajectories using an extended Social Force Model (SFM), reformulated as a system of ordinary differential equations (ODEs). A Physics-Informed Neural Network (PINN) is employed to compute the numerical solution of the ODEs, with an additional loss term introduced based on the residual between the PINN output and the predictions of a GRU model pretrained on real-world data. To represent distinct pedestrian states—such as rushing, yielding, and normal walking—and their subtle transitions, state-specific SFM goal forces are defined, and a continuous soft state-switching mechanism is introduced based on relative position and time-to-collision (TTC). The pretrained GRU captures temporal dependencies from historical trajectories and regularizes the PINN, enhancing the learning of complex behavioral patterns under sparse data conditions. Experiments on the CITR dataset demonstrate that the hybrid model reduces Average Displacement Error by 30.8% and Final Displacement Error by 21.0% compared to a GRU-only baseline, while generating trajectories that align with human commonsense decision-making, achieving high accuracy and physical plausibility even in low-data scenarios.

• Publication year

  2025

• Venue

  International Conference on Service Operations and Logistics, and Informatics

• Publication date

  2025-09-28

• Fields of study

  Not labeled

• Identifiers
  DOI 10.1109/SOLI66289.2025.11396103
• External record

  Open on Semantic Scholar

• Source metadata

  Semantic Scholar

• No claims are published for this paper.

• No concepts are published for this paper.

• Style Nursing with Spatial and Semantic Guidance for Zero-Shot Traffic Scene Style Transfer

  2025cited by this paper
• MGFormer: a multi-information-based GRU-transformer network for pedestrian trajectory prediction

  2025cited by this paper
• Worst Perception Scenario Search via Recurrent Neural Controller and K-Reciprocal Re-Ranking

  2024cited by this paper
• Global-Mapping-Consistency-Constrained Visual-Semantic Embedding for Interpreting Autonomous Perception Models

  2024cited by this paper
• Generating Realistic VRU-AV Scenarios via Social Force-Based Gradient Optimization

  2024cited by this paper
• A CNN-GRU Based Hybrid Approach for Pedestrian Trajectory Prediction

  2024cited by this paper
• Safety-Critical Scenario Generation by Causal Influence Detection

  2024cited by this paper
• Switching: understanding the class-reversed sampling in tail sample memorization

  2022cited by this paper
• Unified regularity measures for sample-wise learning and generalization

  2021cited by this paper
• Worst Perception Scenario Search for Autonomous Driving

  2020cited by this paper
• Social-STGCNN: A Social Spatio-Temporal Graph Convolutional Neural Network for Human Trajectory Prediction

  2020cited by this paper
• A Social Force Based Pedestrian Motion Model Considering Multi-Pedestrian Interaction with a Vehicle

  2020cited by this paper
• A Multi-State Social Force Based Framework for Vehicle-Pedestrian Interaction in Uncontrolled Pedestrian Crossing Scenarios

  2020cited by this paper
• Structure-aware Roadview Synthesis for Testing Autonomous Perception

  2020cited by this paper
• Top-view Trajectories: A Pedestrian Dataset of Vehicle-Crowd Interaction from Controlled Experiments and Crowded Campus

  2019cited by this paper
• Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

  2019cited by this paper
• Social LSTM: Human Trajectory Prediction in Crowded Spaces

  2016cited by this paper
• Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation

  2014cited by this paper
• RoadView: A traffic scene simulator for autonomous vehicle simulation testing

  2014cited by this paper
• Social force model for pedestrian dynamics.

  1995influential reference

• No citing papers are available for this paper.