Visionaid: An Assistive System for Navigation and Environmental Awareness in Visually Impaired Users

Abstract

Blind and visually impaired individuals face persistent barriers to independent travel, financial transactions, and social recognition, and existing assistive aids typically address these needs in isolation while often depending on cloud connectivity. This study designs and evaluates a low-cost, wearable, multi-modal assistive device that consolidates obstacle avoidance, guide-path following, banknote identification, and acquaintance recognition, with all artificial-intelligence inference performed on the device itself. The prototype was built around a Raspberry Pi single-board computer and an OAK-D Lite stereo-depth smart camera, coordinated by the Robot Operating System (ROS 2), and provides three voice-selectable modes: a walking mode combining stereo-depth obstacle detection with colour-based tactile-paving guidance; a banknote mode using an on-device object-detection network for five Thai denominations; and a face mode identifying a pre-enrolled set of individuals. Interaction is fully non-visual, relying on Thai and English speech commands with spoken text-to-speech feedback, and the recognition models were evaluated on labelled image sets. The banknote model achieved a macro mAP@0.5 of 0.966 across five denominations (macro precision 0.972, recall 0.879), with the ฿100 class showing the highest confusion rate; the face module achieved 99.9% accuracy over three enrolled individuals; and obstacle detection reliably flagged objects within 1.5 metres across left, centre, and right zones at 30 frames per second. These results demonstrate the feasibility of an integrated, on-device, voice-driven wearable that delivers multiple independence-supporting functions on affordable embedded hardware.