A Computationally Efficient Model for Pedestrian Motion Prediction

Batkovic, Ivo; Zanon, Mario; Lubbe, Nils; Falcone, Paolo · 2018 · Crossref

DOI: 10.23919/ecc.2018.8550300

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 presents a computationally efficient mathematical model for predicting pedestrian motion over short horizons, specifically designed for collision avoidance algorithms in autonomous driving. The research addresses the challenge of modeling stochastic pedestrian behavior in structured urban environments, such as intersections and crosswalks. While existing methods like Reinforcement Learning (RL) offer accurate long-term predictions, they suffer from high computational complexity, making them unsuitable for real-time vehicle applications. The authors propose a model that assumes rational, road-code-abiding behavior and utilizes a graph-based representation of road geometry to balance accuracy with low computational cost. The proposed method models pedestrian dynamics using a unicycle equation, which is linearized around reference paths defined by the center lines of sidewalks and crosswalks. These references form a connected graph covering all pedestrian areas. The control input is determined using a Linear Quadratic Regulator (LQR) formulation, which penalizes deviations from the reference path. To handle uncertainty, process noise is introduced to predict the covariance of future position deviations. When pedestrians approach junctions or bifurcations, the model propagates predictions along all neighboring edges, effectively branching the trajectory to account for multiple possible directions. This approach avoids the need for expensive offline policy learning required by RL-based methods. The model was validated through simulations and tested against real-world data collected from an intersection in Dresden, Germany, using an infrared camera. The dataset comprised 46 pedestrian trajectories. The authors compared their LQR-based model against a state-of-the-art RL-based model [17]. In terms of accuracy, the LQR model achieved comparable mean position errors over short horizons (up to 10 seconds). However, the LQR model produced more conservative covariance estimates, particularly in symmetric road structures where the RL model struggled with multimodal trajectories. Crucially, the LQR model demonstrated significantly lower computational runtime. Table I shows that for a 200-step horizon, the LQR model required approximately 19.5 milliseconds, whereas the RL model required 4.31 seconds. This substantial difference in speed highlights the LQR model's suitability for real-time implementation. The study concludes that the proposed graph-based LQR model provides a viable solution for short-term pedestrian motion prediction in autonomous driving. It offers accuracy comparable to more complex methods while maintaining the low computational complexity necessary for onboard vehicle processing. The authors note that while the current model is limited to predicting nominal, road-code-abiding behavior, it can be integrated with intent prediction algorithms to handle non-standard behaviors. Future work aims to refine parameter tuning and extend the model to account for situational awareness and non-cooperative agent behaviors.

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.

• anticipation

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Theoretical Contribution: computational model