Modeling Trajectory-level Behaviors using Time Varying Pedestrian Movement Dynamics

Bera, Aniket; Kim, Sujeong; Manocha, Dinesh · 2018 · Crossref

DOI: 10.17815/cd.2018.15

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of modeling realistic, dynamic pedestrian movement in multi-agent simulations. Existing data-driven methods often rely on offline training with fixed parameters or struggle with noisy, sparse trajectory data extracted from videos, failing to capture the time-varying nature of real-world crowd behaviors. The authors propose a novel interactive algorithm that learns Time-Varying Pedestrian Movement Dynamics (TVPMD) directly from 2D trajectories extracted from crowd videos. This approach enables the generation of plausible virtual agent trajectories and improves long-term pedestrian prediction by adapting to dynamic changes in movement patterns on the fly. The method employs statistical techniques to extract three key components of pedestrian dynamics: entry points, movement flows, and local collision avoidance rules. To handle noise and incomplete tracks in video-extracted trajectories, the authors use Bayesian inference, specifically an Ensemble Kalman Filter (EnKF) combined with Expectation Maximization (EM), to estimate the most likely state (position, current velocity, and preferred velocity) of each pedestrian. Movement flows are learned by clustering behavior features—defined by position, average velocity, and preferred velocity over a short time window—using K-means clustering. Entry points are modeled using a multivariate Gaussian Mixture Model (GMM), updated iteratively via EM to reflect recent observations. This allows the system to capture local variations and dynamic shifts in crowd behavior without requiring future knowledge or extensive offline retraining. The authors evaluated their approach on various indoor and outdoor scenarios using real-world videos containing tens of pedestrians. Results demonstrate that the TVPMD framework effectively generates collision-free trajectories for virtual agents that mimic the lane formations and flow patterns observed in original videos. Specifically, the method achieved up to a 12% improvement in long-term pedestrian prediction accuracy compared to baseline methods. Additionally, the use of prior distributions for updating entry point models reduced the average number of EM iterations by more than threefold, significantly improving computational efficiency. The system operates at interactive rates on a desktop PC, with feature computation times reduced by approximately 0.35 seconds for data-driven simulation tasks. The significance of this work lies in its ability to bridge the gap between raw video data and realistic, adaptive crowd simulation. By learning time-varying dynamics rather than static global parameters, the method captures subtle, dynamic aspects of pedestrian motion that previous agent-based models missed. This facilitates robust data-driven crowd simulation in varying environments and enhances trajectory prediction, offering a practical tool for applications in urban planning, robotics, and evacuation planning where realistic crowd behavior is critical.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• pedestrian behavior perception