A hierarchical indoor spatial-semantic reasoning-based scene graph construction for elderly-centric safety warnings

• Enhanced YOLOv7-MLT with multi-scale fusion and attention boosts small-object detection. • Hierarchical spatial-semantic reasoning enables real-time multi-risk prioritization (91.7% accuracy). • Achieves 95.2% fire and 97.6% scald risk detection in 16 indoor elderly-care scenarios. • Dynamic graph embedding evaluates spatiotemporal interactions for adaptive safety warnings. • Integrates scene graphs and cognitive systems for proactive elderly safety monitoring. Indoor safety is crucial for the health and well-being of elderly individuals. However, traditional safety warning systems are hindered by high misdetection rates and delayed alerts, primarily due to their limited ability to perceive multidimensional risks in complex environments and their inadequate reasoning of spatial-semantic relationships. In this study, we propose a hierarchical spatial-semantic reasoning framework for constructing indoor scene graphs, specifically designed for elderly-centric safety monitoring. Our approach utilizes an enhanced YOLOv7-MLT network for dense detection of small objects, incorporating a multi-scale spatially adaptive feature fusion (MSAFM) module and a triplet attention mechanism. These components collaboratively build scene graphs that represent objects, spatial layouts, and associated risk factors. Additionally, graph embedding techniques are employed to extract critical cognitive information, such as element stability and floor slipperiness, while hierarchical warning rules are established to facilitate dynamic risk assessment. This research presents an effective solution for proactive safety monitoring in complex indoor environments.