Publication Details

Abstract

The worldwide prevalence of visual impairment—approximately 2.2 billion individuals according to the World Health Organization—motivates the design of autonomous, affordable assistive wearable systems capable of real-time environmental understanding. Existing mobility tools such as white canes and GPS-enabled navigation applications offer limited semantic comprehension and remain either cost-prohibitive or dependent on internet infrastructure. This paper presents AI-SenseVision, a standalone wearable device that fuses a YOLOv3 deep learning object detection engine with dual ultrasonic proximity sensors to provide instant auditory scene descriptions. The system is trained on a composite dataset of 95 object categories, achieving a mean average precision (mAP) of 83.44% at IoU = 0.5, an F1-score of 86.42%, and obstacle detection accuracy of 95% with end-to-end audio feedback latency under 0.6 s. A multi-instance counting module suppresses repetitive narration, while gesture-based mode selection enables hands-free control across indoor, outdoor, and OCR reading modes. Controlled trials with 47 participants confirm substantial improvements in navigation efficiency, task-completion time, and subjective usability relative to conventional assistive solutions.