Safety-Prioritised Multimodal Navigation for Edge Smart Glasses with Constrained On-Device Language Models

AI-enabled assistive navigation for visually impaired individuals must satisfy strict safety, latency, and computational constraints on wearable edge platforms. This paper presents a safety-prioritised multimodal navigation framework implemented on a fully custom-designed and 3D-printed smart glasses platform integrated with an NVIDIA Jetson Orin Nano edge computing module. The system combines lightweight visual object detection, binaural audio localisation, and QR-based contextual cues with constrained small-scale large language model (LLM) reasoning. A deterministic finite state machine (FSM) enforces a hazard-first decision hierarchy, ensuring that safety-critical detections override contextual navigational suggestions while maintaining real-time operation entirely on-device. Experimental results demonstrate 99% audio event recognition accuracy and up to 86.5% precision and 84.8% recall for object detection under edge constraints. The 0.5B parameter LLM achieves approximately 2.1 seconds inference latency on the Jetson Orin Nano. The integrated system was validated through a structured 9-minute 45-second indoor navigation trial involving dynamic and static obstacles, during which no unsafe actions occurred. These findings show that compact multimodal models, when embedded within a deterministic safety hierarchy and deployed on dedicated edge hardware, enable reliable real-time assistive navigation without reliance on cloud infrastructure.