A Wearable Deep-Learning System for Real-Time Object Detection and Face Recognition to Assist Visually Impaired Navigation

Abstract

The integration of Artificial Intelligence (AI) into wearable assistive technologies offers transformative potential for enhancing mobility, spatial awareness, and autonomy among Blind and Visually Impaired (BVI) individuals. This study introduces a real-time, AI-powered wearable system combining object detection, scene understanding, and facial recognition for safe navigation and environmental interaction. The system comprises smart glasses with an embedded camera and a smart cane equipped with ultrasonic and environmental sensors for detecting obstacles and weather conditions. A lightweight YOLOv8-based model, optimized through hybrid pruning and quantization, ensures efficient inference on resource-limited wearable hardware. Evaluated on MS-COCO 2017, the object-detection module attains an average precision (AP) of 52.9 on the standard COCO 0–100 scale—competitive with substantially heavier two-stage and transformer detectors—while the face-recognition module, evaluated on a custom dataset, reaches 99.3% precision; these two metrics are reported on distinct scales and should not be conflated. The system delivers responsive real-time auditory feedback, guiding users via voice prompts for object localization, obstacle avoidance, and social interaction. This work contributes a scalable, intelligent mobility aid that leverages AI and wearable computing to support independent navigation for BVI users.