Beyond pixels: A comprehensive survey on deep learning-based scene understanding for human-centered applications

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Artificial Intelligence (AI) represents an enabling technology to provide users with adaptive and context-aware support, in line with the Industry 5.0 paradigm. Among AI approaches, Deep Learning (DL) has proven effective in interpreting visual information, which captures environmental features, human activities, and interactions. In this context, high-level scene understanding refers to methods that go beyond object detection or image segmentation, aiming to model interactions, contextual dependencies, and meaningful relationships within visual environments. This capability is essential for supporting people in complex and information-rich scenarios, where low-level visual analysis alone is insufficient. This survey aims to identify the primary contexts where high-level scene understanding supports individuals in their activities, providing context-aware feedback and enhancing safety and situational awareness. Three application areas emerged as most representative: construction environments, where complex and critical scenarios require safety monitoring; surgical environments, characterized by high-risk procedures that demand precise context-aware support; and assistive and inclusive environments for users with special needs, particularly visually impaired individuals, who require a tangible solution for vision support. Across these domains, the review maps domain-specific needs, examines the forms of support via DL technologies, and highlights the constraints that limit real-world adoption. The findings suggest that future progress requires scalable data annotation, the integration of 3D and temporal information, efficient models, and a user-centered, interdisciplinary approach to ensure alignment with actual needs and enhance trust, usability, and real-time applicability.