Designing Visually Accessible Spaces (DEVA)
Overview
Background
Approach
Personnel
Information limited to members of the DEVA
project
Overview
This is a multi-disciplinary project involving personnel from the
University of Minnesota, the University of Utah, and Indiana University,
and supported by
the National Eye Institute of
the National Institutes of Health
grant 1 R01 EY017835-01.
Our long-term goal is to provide tools to enable the design of safe
environments for the mobility of low-vision individuals and to enhance
safety for others, including the elderly, who may need to operate under
low luminance and other visually challenging conditions. We conceive of
a computer-based design tool in which complex, real-world environments
(such as a hotel lobby, large classroom, or hospital reception area),
could be simulated with sufficient accuracy to predict the visibility
of key landmarks or obstacles (e.g., steps or benches) under a variety
of natural and artificial lighting conditions.
Background
Visual accessibility refers to
the use of vision to travel efficiently and safely through an
environment, to perceive the spatial layout of key features in the
environment, and to keep track of one's location in they layout.
There are several million people in the United States with visual
impairments serious enough to restrict their reading and
mobility. Our aim is to explore mechanisms for increasing visual
accessibility for such low
vision individuals.
Low vision comes in many forms, involving combinations of loss of
acuity, contrast sensitivity, and visual field (examples). To construct a
public space that facilitates visual accessibility, it is necessary
predict how well individuals with low vision can perform critical
actions within that space.
One of the
main problems in designing for visual
accessibility arises from the difficulty of predicting the
photometric appearance of real spaces. The image at the far left shows
a
child’s block sitting on top of a flat surface, photographed in
household
lighting which was not manipulated in any way for the purposes of
taking this photograph. Even though there is high contrast in the
reflectivity between one side of the block and the supporting
surfaces, the complex interactions between geometry, materials, and
lighting lead to an image in which there is little visual contrast
between these two surfaces. At the same time, high visual contrast
appears between another side of the block and the supporting surface,
even though the two have identical reflectance values. This is
not just a artificial example. Shown next to the black is a
photograph of a bench that was until recently outside the Psychology
building at the University of Minnesota. Under direct sunlight,
the bench was visible even to those with limited acuity or contrast
sensitivity. Under diffuse lighting such as a cloudy day,
however, the bench is visually indistinct from the walkway and thus
represents a serious hazard for low-vision pedestrians.
While our focus
is on visual accessibility, we are more broadly interested in universal design maximizing the
utility of spaces for all people, regardless of age or
disability. Universal design is a broad, integrated solution to
help everyone, rather than the use of separate solutions for people
with disabilities. It involves the exploitation of key features
important to function,
safety, and mobility. In the case of visual accessibility this
requires an understanding of the complex interactions between geometry,
lighting, and surface properties of objects as they relate to
visual performance.
The image at the left shows the interior of the Gateway Center at the
University of Minnesota, a space in which the conflicts between
aesthetics and the requirements of universal design are quite stark.
Diffuse
lighting, typical of many office environments, avoids strong shadows
and shading effects. As a result, adjacent surfaces with
distinctly different reflectances usually lead to high contrast
edges. In lighting design, aesthetics ("mood") often require the
use of non-diffuse lighting. This generates a conflict between
utility, particularly when low vision is involved, and appearance.
Spot
lighting is
an example of lighting design
intended to facilitate a particular mood in a space. It does so
by generating shadows and strong shading, both of which can impede
visual accessibility in some circumstances. Spot lighting can
also generate disabling glare with many categories of low vision.
In other
circumstances such as the bench shown above, shadows and shading can
facilitate the visual recognition of obstacles better than the visual
contrast between obstacle and background alone.
Architectural
spaces often receive direct or indirect sunlight from windows.
While this often improves the aesthetics of the space and can reduce
energy consumption for lighting, it can also result in shadows and
glare that hurt visual accessibility.
Approach
The project has four closely integrated parts:
- Engineering: Develop
methods for predicting the physical
levels of light reaching the eye for a given viewpoint, geometric
configuration, mix of surface materials, and lighting situation. While
calibrated, high dynamic range (HDR) photometric measurements can be
made of existing spaces, spaces under design will require state-of-the
art computer graphics tools that can reliably predict light levels for
detailed models of physical spaces.
- Empirical: Develop
methods for investigating perceptual
capabilities critical to visually-based mobility, including detection
and classification distances for functionally significant targets in
real environments for a range of lighting conditions and restricted
viewing conditions, and the ability to use these targets for successful
spatial orientation. Utilize these methods to acquire performance data
for normally sighted subjects with visual restrictions and people with
low vision in controlled spaces that are nevertheless representative of
the visual accessibility problems likely to be encountered in real
spaces.
- Computational: Develop
models that can predict perceptual
competence on tasks critical to visually-based mobility, given
photometrically accurate information about a particular visual
environment. Extend these models to account for characteristic forms of
low vision, including severe peripheral field loss, central loss, and
depressed contrast sensitivity. Development of these models will be
informed by the results of empirical testing and will be validated by
testing predicted performance in real-world environments.
- Application: Demonstrate
a proof-of-concept tool that
operates on design models from one or more existing architectural
design systems and is able to highlight potential obstacles to visual
accessibility by predicting the quantitative photometry that would
result from the physical instantiation of the design model and using
this as input to the perceptual model.
Photometrically correct imaging
Predicting the visual accessibility of a space requires knowing the
distribution of light seen from any vantage point of interest. In
computer graphics and photography, relative distributions of light
matter much more than the actual metric values. For this project,
calibrated values are required in order to correctly determine contrast
and account for scotopic
(low light) effects. To capture the variability of lighting that
can occur in real spaces, high
dynamic range (HDR) imaging techniques as representations are
required.
For existing
spaces, multiple exposures can be combined to produce an HDR image
using a variety of tools, including Photoshop,
pfstools,
and several others. For
most purposes, calibration is required to correctly scale values.
This involves specialized light measuring devices and at least some
amount of tedious hand processing. Our main emphasis is in
identifying potential hazards to visual accessibility in the design
process, before new architectural spaces are constructed. This
requires that light intensities be accurately predicted based on design
models. The image at the left was generated by the Galileo
system, developed at the University of Utah. Galileo is capable
of generating photometrically correct images involving complex light
sources, materials models, and participating media.
Empirical testing
Empirical testing will be done to explore obstacle avoidance and
mobility in low vision over a range of environmental spaces, lighting
conditions, and visual deficits. The intent is to balance the
need for experimental controls with the need to evaluate conditions
that are meaningful to the design of real spaces.
For obstacle detection and classification, the first step is to
construct prototypical obstacles in a room in which we can completely
control the lighting. Four classes of obstacles will be
considered: curb-like steps up or down, objects extending up
from the floor, holes in the floor, and objects extending down from the
ceiling (see figure below). Lighting will simulate diffuse
lighting typical of a standard office environment, spot lighting as is
often found in public spaces, and daylight from a side window.
Subjects will include both those with low vision and normal vision
individuals wearing devices intended to simulate low vision. One
open question we hope to contribute to is whether or not using
normal vision subjects wearing visual restrictors provides useful
insights into actual low vision performance.
Two types of mobility tasks will be explored, one involving the ability
to judge distances to points of interest in the environment and the
other involving the ability to maintain a sense of the locations of
points of interest while moving through a space. As with obstacle
detection, subjects will include both those with low vision and normal
vision individuals wearing devices intended to simulate low vision.
Perceptual modeling
The image below illustrates a common low vision problem: How can
obstacles such as curbs be detected in the presence of similar looking
non-hazards and visual clutter?
Although we are still far from a complete model of
human object recognition, there is growing consensus regarding its
overall computational architecture. Evidence from computational,
behavioral, and neural studies suggests the following picture. Visual
recognition begins with a fast feedforward process that extracts
features. These features serve to rapidly
index or propose candidate object or scene categories, such as
“post“, “curb”, “car”, “sign”, etc. Then
depending on the confidence level required for specific task goals,
the decision can be refined by verification through feedback. For
example, there may be sufficient
information in the initial feedforward pass to hypothesize the
existence of an object at an approximate location in the image, but
more iterations and/or more fixations may be required to increase the
confidence of the classification or to accurately localize the object
and determine its precise shape and extent.
Our initial approach to perceptual modeling
will focus on feedforward, image-based theories for several reasons:
1) The first feedforward pass is likely to carry the most critical
information for obstacle avoidance; 2) Image-based methods are
computationally more tractable with
natural image input; 3) There are intriguing
correspondences between image-based approaches and biological vision;
4) It will be straightforward to measure
performance as a function of key input variables characterizing human
low-vision (e.g. loss of spatial resolution, contrast, visual field);
5) The methods we propose are extensible to adaptive perceptual
learning of the important features for a specific task; 6) This
approach extends naturally to using information in a motion sequence
of images; 7) Objects of a particular class can be localized in the
image, providing information for obstacle search; 8) Of particular
relevance to the current project is a fragment-based scheme which
selects fragments that are important areas in the image. Thus
after learning we can use the selected fragments to predict which
features are important for detection. This knowledge could be useful
for designers, since they can focus on improving detectability of
important features.