SafeVision: Vision-language reasoning for context-aware safety monitoring

초록

Safety compliance in high-risk industrial domains such as construction and mining remains a persistent challenge due to the complex interaction of workers, heavy machinery, and complex site conditions. Conventional com puter vision methods can detect objects or classify activities but often lack the semantic reasoning needed for comprehensive safety assessment. This paper presents SafeVision, a unified vision-language framework for realtime, context-aware safety assessment. SafeVision integrates a frozen CLIP-ViT encoder with two complementary reasoning modules: the Scene-Aware Reasoner (SAR) for global environmental context understanding and Region-Focused Neural Reasoner (ReFiNER) for region-level safety compliance verification. A task-conditioned visual question answering component further enables natural language queries regarding hazards, personal protective equipment (PPE) usage, and operational conditions. To support deployment on resource-constrained platforms, SafeVision employs parameter-efficient Low-Rank Adaptation (LoRA), ensuring scalability without loss of in terpretability. Experiments on a custom multimodal dataset of 10,000 annotated images demonstrate strong performance across multiple tasks, with SAR reaching 87.4 % accuracy in scene classification and the VQA com ponent achieving BLEU-4 of 83.2 %, F1 of 84.6 %, Exact Match of 80.3 %, and METEOR of 82.5 %. Qualitative findings illustrate effective alignment between global scene understanding and localized reasoning, underscor ing the potential of SafeVision for proactive safety monitoring in industrial environments. The source code and dataset are available at: https://github.com/Murtazaabidi1/SafeVision-Vision-Language-Reasoning-for-Context-Aware-Safety-Monitoring.

키워드