NFPA Journal - July/August 2020 - 61

NFPA

While poor management of crowded
events routinely leads to civilian injuries and deaths, modern tools and
technologies can be used to enhance
crowd management strategies. Existing
technologies and simulations designed
to assess crowd densities and dynamics can be useful, but they are often
expensive and challenging to implement, with only questionable accuracy
that provides little assistance to crowd
managers.
The NFPA web-based crowd monitoring tool, by contrast, is focused
on overcoming these challenges, and
utilizes modern technologies and algorithms to evaluate crowd movement
over time in high-risk spaces. The tool

can be accessed with a laptop, tablet,
or mobile phone. Users would simply
upload a still image or a short video
clip to the site, or connect to a streaming video feed. For still images, the tool
can run an evaluation for crowd count
in a given area. For video options, the
tool can capture individual frames
at a user-specified interval, ranging
from one to 30 seconds, or manually
as desired. The frame capture is then
evaluated by the model, which predicts the count and provides a crowd
density map, which is rendered for the
user in about five seconds. The counts
are then graphed at the timestamp of
each capture, based on the user-specified interval, and displayed to identify

p The crowd management tool under development by NFPA can analyze a photograph
or video clip of a crowd and generate a heat map illustrating the areas of highest density.
This real-time analysis can help safety officials make potentially live-saving decisions.

trends in crowd density. Early trials
have been very promising; the tool
has been accurate when subjects were
either in the foreground or in the background-an important improvement
over previous models, which tended to
search for images of complete individuals and had difficulty identifying people
who were obscured.
Our tool is based on a concept
known as computer vision, a subset of
computer science that enables us to
teach computers to recognize objects
in a given scene. Within the computer
vision and deep learning communities,
this process is defined as congested
scene recognition, or CSR, which has
two primary objectives: counting the
number of individuals in a crowd, and
identifying their spatial distribution.
The second part is especially important
to facility managers, since two different
scenes containing the same number
of people can have vastly different
distributions.
We are testing the feasibility and
functionality of our initial prototype,
including the accuracy of crowd estimations, timeliness of reporting, value
of information, ease of implementation
and use, economic feasibility, and computational demand. Our hope is that
this tool puts data-informed crowd
management within reach of safety officials who have found previous versions
to be expensive, inaccurate, or otherwise problematic. That's in part why we
intentionally built the tool using opensource platforms and general-purpose
programming languages, where the
source code is freely available.
The risks associated with large
crowds in a variety of venues will
never go away, and crowd managers
will always face immense responsibilities that cannot be replaced by
technology alone. But it is our hope
that new tools and technologies can
give crowd managers the relevant,
real-time information they need to
make split-second decisions and keep
people safe.
VICTORIA HUTCHISON Victoria Hutchison
is a research project manager with the Fire
Protection Research Foundation. FREDERICK
MACDONALD, a data scientist at NFPA,
contributed to this article.

N F PA . O R G / J O U R N A L * NFPA JOURNAL

Crowd Control feature_08.20 R2 DCS FINAL.indd 61

| 61

7/1/20 5:57 PM

http://NFPA.ORG/JOURNAL

NFPA Journal - July/August 2020

Table of Contents for the Digital Edition of NFPA Journal - July/August 2020