Abstract:
Navigating from one place to another has been a problematic task for the blind. In
Bangladesh, the existing footpaths are mostly crowded or broken. Often, visually impaired
people get hurt while walking on a footpath as they do not have anything but a stick to help
them. Considering the problem scenario, we are proposing a smart solution to identify safe
footpath and detect obstacles in a footpath. The system will also be capable of estimating
the distance of the object as well as suggesting the safe pathway. To train the models we
built a dataset of footpath images of Dhaka containing 3,000 hand-annotated RGB images
for semantic segmentation and another dataset containing 500+ samples of real-world
distances of reference objects w.r.t to their pixel coordinates in an image for distance
estimation. We adopted and modified the U-Net architecture that is trained on our
segmentation dataset which is capable of inference safe footpath with 96% accuracy with
as low as 4.7 million parameters. The system utilizes YOLOv3 architecture for object
detection and a polynomial regression based novel approach to estimate object distance.
The distance measurement model obtains a score of 94%.