A synthetic-vision based steering approach for crowd simulation

Ondřej, Jan; Pettré, Julien; Olivier, Anne‐Hélène; Donikian, Stéphane · 2010 · OpenAlex-citations

DOI: 10.1145/1778765.1778860

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 generating realistic, collision-free navigation for crowds of virtual walkers in computer animation. While existing agent-based techniques often rely on explicit geometric models to compute admissible velocities, these approaches can disconnect from human behavior and struggle with complex interactions, leading to artifacts or unrealistic motions. Motivated by cognitive science research on human locomotion, the authors propose a novel synthetic-vision-based steering approach. The core premise is that humans achieve safe navigation by extracting specific information from their optic flow rather than performing explicit geometric collision checks. The goal is to design a local collision avoidance method that mimics this visual-stimuli/motor-response control loop, thereby enhancing the believability of animations and promoting the emergence of self-organized patterns seen in real crowds. The method is grounded in the work of Cutting et al., which posits that humans detect future collisions by monitoring the time derivative of the bearing angle ($\dot{\alpha}$) of perceived obstacles and assess danger via time-to-collision (ttc). The model treats each walker as having a synthetic vision system that perceives static and moving obstacles as points. For each perceived point, the system computes the bearing angle, its derivative, and the time-to-interaction. The motor response is twofold: a reorientation strategy avoids future collisions by adjusting angular velocity based on $\dot{\alpha}$, while a deceleration strategy handles imminent collisions by reducing tangential velocity when ttc falls below a threshold. The implementation utilizes OpenGL and CUDA to efficiently compute these visual inputs. Obstacles are simplified into geometric primitives (e.g., cones for other walkers) to optimize performance, and the visual data is processed via shaders to determine the necessary steering adjustments without explicitly calculating admissible velocity domains. The results demonstrate that this vision-based approach successfully generates emergent, self-organized patterns in crowd simulations, such as lane formation and passing behaviors, which are visually appealing and consistent with real human interactions. Compared to previous geometric and rule-based approaches, the model improves the global efficiency of walker traffic and significantly reduces improbable locking situations where agents get stuck. The simulations show that walkers can navigate complex environments with numerous obstacles and other moving agents smoothly. By relying on visual cues rather than explicit collision checks, the method avoids the computational pitfalls of degenerating admissible velocity domains in complex multi-agent interactions. The significance of this work lies in its integration of cognitive science principles into crowd simulation, offering a more biologically plausible model for locomotion control. It challenges the dominance of purely geometric avoidance models by demonstrating that simpler perception-action loops can yield robust and realistic results. The approach enhances the overall realism of crowd animations by ensuring that local interactions naturally lead to coherent global formations. Furthermore, it provides a framework for handling complex interactions implicitly through visual projection and filtering, addressing limitations in previous methods regarding the combination of multiple interaction rules. This contributes to the field by bridging the gap between computational efficiency and behavioral realism in virtual crowd simulation.

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.

• perception action locomotion