Context-Aware Vision–Language Safety Reasoning for UxVs = Context-Aware Vision–Language Safety Reasoning for UxVs

다국어 초록 (Multilingual Abstract)

This thesis addresses real-time safety reasoning for unmanned ground and aerial vehicles (UxVs) operating on resource-constrained edge platforms. Practical field operation requires not only robust perception in challenging scenes (low light, clutter, occlusion) but also human-understandable explanations that support operator trust and rapid decision making. To meet these needs, this thesis presents SafeVision, a parameter-efficient vision language system that delivers scene-level understanding, region-grounded safety checks, and visual question answering. SafeVision couples a froze CLIP-ViT visual encoder with two complementary reasoning paths: a Scene-Aware Reasoner (SAR) that captures global context and a Region Focused Neural Reasoner (ReFiNER) that localizes PPE/hazard evidence. A lightweight cross-modal adapter and task tokens ([scene], [region], [vqa]) condition a LoRA-tuned language model to produce grounded, interpretable outputs while keeping trainable parameters small for edge deployment. The approach is evaluated on a custom industrial-safety dataset with scene, region, and VQA annotations. Empirically, SAR achieves 87.4% scene accuracy, and the VQA head attains BLEU-4 83.2, F1 84.6, EM 80.3, and METEOR 82.5, with qualitative examples demonstrating clear alignment between visual evidence and textual reasoning. Prototype deployments on UAV/UGV platforms show that SafeVision meets typical on-board latency and memory budgets.