Microscopic crowd simulation : evaluation and development of algorithms

David Wolinski 1, 2
1 Lagadic - Visual servoing in robotics, computer vision, and augmented reality
CRISAM - Inria Sophia Antipolis - Méditerranée , Inria Rennes – Bretagne Atlantique , IRISA-D5 - SIGNAUX ET IMAGES NUMÉRIQUES, ROBOTIQUE
2 MIMETIC - Analysis-Synthesis Approach for Virtual Human Simulation
IRISA-D6 - MEDIA ET INTERACTIONS, Inria Rennes – Bretagne Atlantique , UR2 - Université de Rennes 2
Abstract : With the considerable attention crowd simulation has received, many algorithms have been and are being proposed. Yet, (1) there exists no standard scheme to evaluate the accuracy and flexibility of these algorithms, and (2) even the most recent algorithms produce noticeable simulation artifacts. Addressing the first issue, we propose a framework aiming to provide an objective and fair evaluation of the realism of crowd simulation algorithms. ''Objective'' here means the use of various metrics quantifying the similarity between simulations and ground-truth data acquired with real pedestrians. ''Fair'' here means the use of parameter estimation to automatically tune the tested algorithms to match the ground-truth data as closely as possible (with respect to the metrics), effectively allowing to compare algorithms at the best of their capability. We also explore how this process can increase a user's control on the simulation while reducing the amount of necessary intervention. Addressing the second issue, we propose a new collision-avoidance algorithm. Where current algorithms predict collisions by linearly extrapolating agents' trajectories, we better predict agents' future motions in a probabilistic, non-linear way, taking into account environment layout, agent's past trajectories and interactions with other obstacles among other cues. Resulting simulations do away with common artifacts such as: slowdowns and visually erroneous agent agglutinations, unnatural oscillation motions, or exaggerated/last-minute/false-positive avoidance manoeuvres. In a third contribution, we also explore how evaluation and parameter estimation can be used as part of wider systems. First, we apply it to insect simulation, taking care of local insect behavior. After completing it at the intermediate and global levels, the resulting data-driven system is able to correctly simulate insect swarms. Second, we apply our work to pedestrian tracking, constructing a ''meta-algorithm'', more accurately computing motion priors for a particle-filter-based tracker, outperforming existing systems.
Complete list of metadatas

Cited literature [81 references]  Display  Hide  Download

Contributor : Abes Star <>
Submitted on : Tuesday, December 20, 2016 - 12:06:07 PM
Last modification on : Friday, January 11, 2019 - 2:28:30 PM
Long-term archiving on : Tuesday, March 21, 2017 - 2:02:25 PM


Version validated by the jury (STAR)


  • HAL Id : tel-01420105, version 1


David Wolinski. Microscopic crowd simulation : evaluation and development of algorithms. Data Structures and Algorithms [cs.DS]. Université Rennes 1, 2016. English. ⟨NNT : 2016REN1S036⟩. ⟨tel-01420105⟩



Record views


Files downloads