CHS: Small: A Perceptual-based Approach to improve Synthetic Crowds

点击翻译按钮获取中文摘要

In principle, crowd simulations could help designers make buildings both nicer and safer by allowing them to explore how people would move through the building in both normal and emergency situations. However, it is unclear whether existing tools for simulating crowds reflect the motion of actual people, which makes designers hesitant to use them. This project's goal is to develop a method that allows people to visually compare the realism of computer-generated crowd movement to real crowd movement. The idea is that the more a simulated crowd looks as real as an actual crowd, then the more likely that the algorithm used to generate the simulation and the simulation itself will be useful in building design and will be trusted by building designers. The team will address these questions by (1) processing videos of real crowd motion in buildings to generate 3D reconstructions, then (2) asking both non-experts and facilities managers to (a) rate the realism of parallel videos that use the real crowd paths versus simulated crowd paths through the reconstructed building and (b) comment on aspects of the videos that affect their ratings. These data will be useful for both evaluating and improving the quality of future crowd simulation algorithms; to this end, the team will release the videos, datasets, algorithms, experimental tools, and results to help other researchers in this and related areas. They will also use the materials in college courses aimed at simulation and modeling, as well as developing outreach experiences for middle and high school students and outreach materials for building designers.To generate the experimental materials, the team will first process crowd videos drawn from existing crowd movement video databases when possible and captured by the team when needed to represent conditions not available in those databases. They will then extract people's initial locations and paths, using Catmull-Rom splines to compensate for noise in the extraction of location and building features and variations in frame rates between videos, as well as creating a 3D reconstruction of the facility. For generating simulated paths, they will use the captured starting location, time, and ending location as input to a suite of open source agent-steering algorithms that represent a wide variety of simulation approaches. To remove visual cues from the original films that might influence judgments, both the original paths and the simulated paths will be generated using the Unity 3D graphics rendering and physics engine, which is commonly used in crowd simulation research. Finally, the team will develop an interface for rating and annotating pairs of videos (and, for longer videos, video segments) for realism. This interface will be used in the experiments described above in which participants compare real and simulated renderings, as well as comparisons between different segments of the same video, to collect as wide a variety of data as possible on factors that affect realism judgments.