Assistive tech — MDes, UC Berkeley · 2021
The task I set myself was to make the world safer and more accessible for the visually impaired, utilising deep-learning image recognition and text-to-voice warnings. The system analyses the scene in front of the user and informs them of objects coming towards them and obstacles they are approaching.
The design adopts an item of daily use the blind already wear — the blacked-out glasses that protect their eyes from sunburn, which causes severe pain even when it cannot further their blindness — so the user does not need to get used to a new wearable.
A key consideration was making the image recognition fully local, using edge architecture: TensorFlow with a localised COCO dataset, so no connectivity problem could endanger the user. A photocell detects when the camera needs additional light for recognition to continue, provided by a NeoPixel ring mounted around the camera.
