Post-Fall Non-Recovery State Detection Using Pose Estimation and Temporal Persistence

This paper presents an edge-based system for detecting post-fall non-recovery states in a predefined indoor monitoring area. The target of the proposed system is not the instantaneous fall event itself, but a fall-like abnormal posture that persists without recovery. Human keypoints and a person bounding box are extracted using a pretrained YOLO26n-Pose model. Torso orientation is used as the primary posture feature, and the bounding-box width-to-height ratio is used as an auxiliary horizontal-posture feature. A finite-state machine separates frame-level abnormal-posture candidates from emergency decisions by assigning Normal, Warning, and Emergency states according to the accumulated posture duration. In the self-recorded scenario evaluation, no Emergency transition was observed in the normal-posture and daily-motion scenarios, whereas sustained abnormal postures produced sequential Warning-to-Emergency transitions. The average pose inference time on the Jetson AGX Orin was 53.12 ms/frame. The UR Fall Detection Dataset, consisting of 30 fall and 40 ADL RGB clips, was additionally used to examine the external applicability of the posture features and their threshold sensitivity. Under the conservative sensitivity setting, abnormal-posture candidates were detected in 90.0% of the fall clips, while the moderate Warning setting achieved a clip-level F1-score of 0.6552. These results should be interpreted as a feasibility and sensitivity analysis under the evaluated conditions rather than as a general performance validation for unrestricted fall-detection environments.