PerceptiCane: Low-Cost Embedded Vision with Lightweight Hazard Ranking for Assistive Mobility

PerceptiCane is a low-cost embedded computer-vision prototype designed to investigate contextual object prioritization for assistive mobility. The system combines YOLO26n object detection with a lightweight hazard-ranking heuristic based on apparent bounding-box size and horizontal alignment within the camera view. The YOLO26n detector achieved a validation-set mAP50 of 58.2% and an mAP50–95 of approximately 0.39 across seven object classes. The hazard-ranking heuristic was evaluated independently using annotated bounding boxes from 2,250 multi-object scenes containing 19,330 objects, with an average ranking computation time of 0.1083 milliseconds per scene. Prototype testing on a Raspberry Pi 4 reported an average inference time of 87 milliseconds per frame, a processing rate of 2.8 FPS, and memory consumption of 340 MB. The findings demonstrate the computational feasibility of combining embedded object detection with lightweight object ranking. However, the ranking heuristic has not been validated against independent human judgments of navigational relevance, and the prototype has not undergone controlled testing with visually impaired participants. Consequently, the results do not establish real-world navigational effectiveness or safety. Publication Status: Research preprint; not peer-reviewed. Generative AI Disclosure: ChatGPT (OpenAI) assisted with manuscript drafting and editing, organization, technical explanations, literature-search support, and verification of experimental calculations. The underlying research and experiments were conducted by the author, who retains responsibility for the scientific content, interpretations, and conclusions.