Interpretable fall detection via scene graphs: decision fusion with constrained LinUCB [Sahne grafi tabanlı yorumlanabilir düşme tespiti: kısıtlı LinUCB ile karar birleştirme]

Official URL: https://dx.doi.org/10.1109/SIU71813.2026.11636824

Falls are a leading cause of serious injuries among older adults; however, vision-based fall detection systems often produce high false alarms due to confusion between support surfaces such as beds or chairs and the floor. This study presents an interpretable fall detection framework in which each video frame is represented as a scene graph encoding person posture, objects, and person-support surface relations. The perception stage integrates object detection, human pose estimation, monocular depth estimation, and segmentation; floor geometry is estimated via RANSAC-based plane fitting, and the most plausible support surface for each person is selected using a scoring function. Changes between consecutive scene graphs are reduced to a 13-dimensional feature vector and evaluated using a constrained LinUCB decision module. Experiments on the URFall and CAUCAFall datasets achieve 0.960 recall and 0.800 F1-score, with alarms reported through humanreadable short explanations.