Satinder Singh won the Jetson Project of the Month for DeepWay, an AI-based navigation assistance system for the visually impaired. The project, which runs on an NVIDIA Jetson Nano Developer Kit, monitors the path of a person and provides guidance to walk on the correct side and avoid any oncoming pedestrians.
In addition to the Jetson Nano, Satinder’s setup includes an Arduino Nano, a webcam, two servo motors, a USB audio adapter, 3D printed eye-glasses and a few peripherals. The Arduino Nano is used to control the servo motors, which nudge the user in the correct direction by gently tapping them on either side of the head. If the Jetson Nano identifies pedestrians, a voice prompt sent through the USB audio adapter can be used to warn the user.
To train his convolutional neural network, Satinder collected 10,000 images across three classes for left, center and right positions within a lane, and trained his model on a GPU-enabled virtual machine on Microsoft Azure. He used both PyTorch and Keras for training a U-net semantic segmentation model. After further analysis, he picked the U-Net model trained in Keras for its better performance.
DeepWay – Project Walk Through
Satinder believes that his solution stands out among others because of its portability and affordability (< $200 USD). Using the sense of touch to alert users of objects or people also helps sight-impaired individuals to keep their hearing, which is essential for daily navigation and safety. He plans to improve the project and has laid out a plan including further optimizing it using NVIDIA TensorRT. We can’t wait to see how this project evolves.
The World Health Organization estimates at least 2.2 billion people around the world have a vision impairment. Many among this group would benefit from navigation aids. Open-source prototypes such as DeepWay foster innovation around these use-cases and are great examples of AI being used for social good.
If you’re interested in building your own ‘DeepWay’ or extending it, Satinder has shared the instructions and the code here.