AI VISION ASSISTANT: A SAFETY-AWARE DESIGN FRAMEWORK AND EVALUATION PROTOCOL FOR OBJECT IDENTIFICATION, TEXT READING, SCENE DESCRIPTION, AND OBSTACLE AWARENESS
Rizwan Shaikh
Abstract
This paper presents AIVA, a safety-aware design framework and evaluation protocol for a smartphone-based visual assistant intended to support blind and low-vision people with object identification, OCR-based text and signboard reading, scene description, and cautious obstacle awareness. The contribution is deliberately framed as a design specification—not as a validated assistive product. Existing work has already integrated combinations of object detection, OCR, audio feedback, scene understanding, and wearable or mobile deployment. Accordingly, the paper does not claim novelty in merely combining these functions. Its proposed contribution is a set of explicit integration rules: user-intent task routing; separation of descriptive scene narration from mobility-related alerts; evidence-linked grounding; uncertainty and abstention behavior; priority-based spoken output; and an evaluation protocol with hypotheses and failure criteria. Candidate data sources, feature definitions, threshold-selection procedures, and fallback behavior are specified for obstacle-risk estimation, while all thresholds remain subject to empirical calibration and safety review. A structured comparison identifies overlap with prior systems and distinguishes the proposed evaluation focus. No prototype measurements, user-study findings, or safety validation are reported. This manuscript is therefore a methodological blueprint, not evidence of improved navigation, accuracy, usability, or safety.