Social Lstm With Dynamic Occupancy Modeling for Realistic Pedestrian Trajectory Prediction
Year: 2026
Type: software
Abstract: In dynamic and crowded environments, realistic pedestrian trajectory prediction remains challenging due to the complex nature of human motion and mutual influences among individuals. Deep learning models have achieved promising results by implicitly learning such patterns from 2D trajectory data. However, most approaches treat pedestrians as point entities, ignoring the physical space each person occupies. To address these limitations, this study proposes a novel deep learning model that enhances Social LSTM with a new Dynamic Occupied Space (DOS) loss function. This loss guides Social LSTM to reduce collisions without increasing displacement error across different crowd densities in both homogeneous and heterogeneous settings. It combines the average displacement error with a new collision penalty sensitive to scene density and individual spatial occupancy. For training and evaluation, five datasets were generated from real pedestrian trajectories recorded during the Festival of Lights in Lyon in 2022. Four datasets represent homogeneous crowd conditions—low, medium, high, and very high density—while the fifth represents a heterogeneous density distribution. The experimental findings show that the proposed model achieves an average relative reduction of approximately 31.5% in collision rate across all datasets, ranging from 7.6% to 57.9%. The model also reduces the average displacement error (ADE) and final displacement error (FDE) by an average of 5.2% and 6%, respectively, compared to the baseline. Against Vanilla LSTM and Social GAN, the proposed model achieves lower collision rates across all five datasets and comparable or better displacement accuracy in four of them. Software and Data Availability This Zenodo repository provides four Python files associated with the proposed DOS loss function and its training and evaluation within the TrajNet++ framework: loss.py, trainer.py, fast_evaluator.py, and metrics.py. A README.md file provides installation instructions, file locations, and usage information. The implementation builds upon the TrajNet++ Baselines and TrajNet++ Tools repositories. The original pedestrian trajectory data are publicly available through the MADRAS data platform. TrajNet++ Baselines: https://github.com/vita-epfl/trajnetplusplusbaselines TrajNet++ Tools: https://github.com/vita-epfl/trajnetplusplustools MADRAS Dataset: https://madras-data-app.streamlit.app/
Authors: Alia, Ahmed, Seyfried, Armin, Chraibi, Mohcine
Institutions: Forschungszentrum Jülich, An-Najah National University
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7223706011