AI VISION ASSISTANT: AN INTELLIGENT ASSISTIVE APPLICATION FOR VISUALLY IMPAIRED PEOPLE USING OBJECT RECOGNITION, TEXT READING, OBSTACLE DETECTION, AND VOICE-BASED SCENE DESCRIPTION
Rizwan Shaikh
Abstract
Visually impaired people frequently encounter barriers when interpreting unfamiliar environments, locating everyday objects, reading printed information and signboards, and identifying potential obstacles during mobility. Existing assistive applications commonly address one or a subset of these tasks, while the practical use of multiple independent tools can create interaction overhead and cognitive burden. This paper presents a modular AI Vision Assistant (AIVA), a smartphone-oriented assistive system design that integrates object detection, scene understanding, optical character recognition (OCR), obstacle-risk estimation, voice interaction, and text-to-speech (TTS) into a unified workflow. The proposed architecture combines a camera-based perception layer with task-specific computer vision modules, a context and priority manager, a grounded language generation layer, and an accessible audio interface. A key design principle is to separate safety-critical perception from descriptive language generation: object and obstacle signals are processed using explicit confidence and risk rules, while a vision-language model is constrained to describe only evidence available from the current image and structured detections. The system supports four interaction modes—object query, text/signboard reading, scene description, and obstacle alert—and uses concise, interruptible speech to avoid overwhelming the user. The manuscript contributes a requirements-driven architecture, a task-routing strategy, a risk-aware audio policy, a privacy-preserving deployment model, and a reproducible evaluation protocol. As no user study or prototype measurements are supplied for this manuscript, performance is not represented as experimentally validated. Instead, measurable hypotheses and evaluation procedures are specified for detection quality, OCR accuracy, latency, speech intelligibility, false-alert burden, and usability with blind and low-vision participants. The proposed design aims to provide a practical foundation for future implementation and user-centered validation of multimodal visual assistance.