Authors :
Akshay Anand
Volume/Issue :
Volume 8 - 2023, Issue 11 - November
Google Scholar :
https://tinyurl.com/4kn3t63y
Scribd :
https://tinyurl.com/2p9662xe
DOI :
https://doi.org/10.5281/zenodo.10255025
Abstract :
Navigating indoor environments can be
challenging for visually impaired people, particularly for
wayfinding tasks. With tools like GPS, outdoor
navigation is more feasible, however, when indoors,
receiving low-precision location data and avoiding
obscure obstacles pose a challenge. We propose an app
that combines state-of-the-art advances in promptable
image segmentation from computer vision and
augmented reality to assist the visually impaired in
indoor navigation. Due to a broader range of objects
indoors, automatically detecting obstacles in real-time is
challenging. The key idea in our approach is to use a
faster variation of Meta’s Segment Anything Model
(FastSAM) to segment objects in the user’s path. We use
a generic indoor map of the environment to localize the
user’s position and overlay AR arrows that guide their
navigation. FastSAM’s zero-shot recognition capabilities
allow us to automatically add nearby obstacles in real-
time to the indoor map so the wayfinding can be updated
to avoid these. Although FastSAM’s speed enables our
app to be deployable in real-time, the performance
tradeoff from the original model makes mask generation
less precise. Overall, our app can detect larger obstacles,
such as chairs and tables, at a high rate and generates
optimal paths to reach a destination. Many existing
indoor navigation systems highly depend on a detailed
indoor map or an extensive 3D environment model and
don’t account for dynamic obstacles. This system
minimizes the amount of initial data needed and can
account for obstacles that cannot be observed from a
map.
Navigating indoor environments can be
challenging for visually impaired people, particularly for
wayfinding tasks. With tools like GPS, outdoor
navigation is more feasible, however, when indoors,
receiving low-precision location data and avoiding
obscure obstacles pose a challenge. We propose an app
that combines state-of-the-art advances in promptable
image segmentation from computer vision and
augmented reality to assist the visually impaired in
indoor navigation. Due to a broader range of objects
indoors, automatically detecting obstacles in real-time is
challenging. The key idea in our approach is to use a
faster variation of Meta’s Segment Anything Model
(FastSAM) to segment objects in the user’s path. We use
a generic indoor map of the environment to localize the
user’s position and overlay AR arrows that guide their
navigation. FastSAM’s zero-shot recognition capabilities
allow us to automatically add nearby obstacles in real-
time to the indoor map so the wayfinding can be updated
to avoid these. Although FastSAM’s speed enables our
app to be deployable in real-time, the performance
tradeoff from the original model makes mask generation
less precise. Overall, our app can detect larger obstacles,
such as chairs and tables, at a high rate and generates
optimal paths to reach a destination. Many existing
indoor navigation systems highly depend on a detailed
indoor map or an extensive 3D environment model and
don’t account for dynamic obstacles. This system
minimizes the amount of initial data needed and can
account for obstacles that cannot be observed from a
map.