Vision-Aid AI Powered Smart Spectacles Indoor Navigation via Custom YOLOv8 Optimization

Independent indoor navigation presents significant spatial awareness challenges for visually impaired individuals. This paper presents a wearable edge-computing assistive system consisting of smart spectacles integrated with a Raspberry Pi 4, a USB camera, an HC-SR04 ultrasonic sensor, and an offline text-to-speech module. To improve indoor object recognition, a custom hybrid dataset containing nine classes was developed by combining selected subsets from the SUN RGB-D and COCO 2017 datasets, focusing on common indoor navigation obstacles. A custom-trained YOLOv8n model performs real-time object detection, while a sequential processing pipeline executes image acquisition, object detection, distance measurement, and speech synthesis using the offline pyttsx3 engine. Experimental evaluation achieved an overall mean Average Precision (mAP50) of 0.577 across the nine object classes, demonstrating the feasibility of deploying an efficient, low-cost indoor navigation system on resource-constrained edge hardware without requiring cloud connectivity. This preprint is made available for early dissemination ahead of formal peer-reviewed submission. The complete implementation, including the training pipeline and Raspberry Pi deployment scripts, is available in the companion code repository (see Related Works, released under Apache License 2.0). The custom hybrid training dataset was derived from subsets of the COCO 2017 and SUN RGB-D datasets. Due to the licensing terms of these source datasets, raw source imagery is not redistributed here; dataset construction and licensing details are documented in the code repository.