Deep Learning-Based Real-Time Object Recognition and Navigation Assistance System for Visually Impaired Individuals

Abstract

Visually impaired individuals face significant daily challenges in object recognition, navigation, and environmental awareness, often relying on canes, guide dogs, or human assistance that limit independence. Traditional assistive technologies like ultrasonic sensors provide basic obstacle detection but lack semantic understanding of surroundings. This paper presents a Deep Learning-Based Real-Time Object Recognition and Navigation Assistance System that uses YOLOv8 object detection and lightweight pose estimation running on smartphone hardware to identify 80+ everyday objects, detect obstacles, read text, and provide spoken navigation guidance. The system processes camera frames at 25 FPS on mid-range Android devices, achieving 92.4% [email protected] detection accuracy and 87.6% navigation success rate in real-world tests across indoor/outdoor environments. Key innovations include edge-optimized model quantization, multimodal output (speech + haptic feedback), and context-aware descriptions that combine object identity, distance, and spatial relationships ("chair 1.2m ahead left, table 0.8m front"). Field trials with 15 visually impaired participants demonstrated 78% reduction in navigation time and 65% increase in obstacle avoidance confidence compared to white cane use alone. The system operates offline with <200MB storage, making it accessible in low-connectivity regions and supporting multiple languages through integrated TTS. These results demonstrate practical deep learning deployment for assistive technology, enabling greater independence and safety for the visually impaired.