From Perception to Cognition: A Systematic Review of Informatics-Driven Vision-Based Safety Management for Sustainable Development in High-Risk Industries

High-risk industries (HRI) face persistent safety challenges due to complex environments and multi-factor risks. Traditional manual monitoring is inefficient and reactive. While computer vision (CV) and deep learning (DL) have enabled automated risk perception, existing research lacks systematic review of the transition toward risk cognition. This systematic review, following PRISMA 2020 guidelines, searched Web of Science, Scopus, and IEEE Xplore for studies published from January 2021 to December 2025. After two-stage screening, 108 eligible studies were included. These studies span construction, mining, oil and gas, petrochemical, rail, energy, and maritime industries, with perception-layer applications predominating while cognition and early warning layer studies remain limited. We propose a “perception–cognition–early warning” framework to map the evolution from data to information, knowledge, and actionable decisions. Our findings reveal that visual perception technologies (e.g., Personal Protective Equipment (PPE) detection, object tracking) have matured, but significant bottlenecks persist in multimodal information fusion, knowledge reasoning, and decision-making. Achieving a true cognitive leap requires multimodal semantic alignment, scene graph construction, and causal inference. The proposed framework enables practitioners to design cognitive safety vision systems with scene understanding and interpretable decision-making capabilities. Key implementation strategies include addressing challenges in few-shot learning, cross-domain generalization, and human–machine collaboration. By integrating AI-driven perception, cognitive reasoning, and proactive intervention, this review supports the United Nations Sustainable Development Goals. Relevant goals include Goal 3 (health and well-being), Goal 8 (decent work), and Goal 9 (industry and innovation).