Trajeform trains a VAE on real crowd trajectories and uses it to check how believable crowd simulators are.
Method: A reconstruction VAE learns a continuous embedding from real trajectory data that keeps the behavioural traits of real crowds. The trained encoder is then applied to simulator output, and the embedding distributions are compared.
Simulators tested: CCP, modCCP and RVO, each compared against ground-truth datasets.
Extras: Supplementary data, such as cycling trajectories, checks how well each simulator's principal behaviours span the real embedding space.
Evaluation: Latent-space analysis, distributional metrics, and trajectory sampling from the aggregated posterior. The paper is 14 pages with 17 figures and 5 tables.
Sign in to comment.
TL;DR of the paper, generated automatically.
Trajeform trains a VAE on real crowd trajectories and uses it to check how believable crowd simulators are.