Cross-Dataset Fall-Event Detection via Causal RGB–Pose Temporal Modeling and Alarm Verification

Falls require rapid assistance, but fall-like activities and domain shift complicate alarm control. We present a cross-dataset fall-detection framework combining red–green–blue (RGB) appearance, pose evidence and sequential decisions. A frozen YOLO26n-Pose (You Only Look Once) front end supports 95-dimensional pose/kinematic and 256-dimensional person-centered appearance features processed by independent causal temporal convolutional networks over 16-frame windows. A quality gate offers auditable fusion without consistent superiority over fixed logits. Three-state outputs generate alarms through causal evidence accumulation, confirmation and logistic verification. CAUCAFall supports source training, URFD/Le2i external development, and GMDCSA24, a one-time blind test after weights, policies and scripts are hash-locked. Historical verification and rescue reduced false alarms from 137.10 to 53.32/h while event recall decreased from 92.41% to 75.95% (60/79 events), with a 1.25 s median delay. Corrected Le2i labels and an eight-feature verifier without rescue retrospectively matched these aggregate counts; this is not a new blind test. Four blind-test subjects, 0.263 h of negative exposure, staged falls and offline alignment restrict conclusions to benchmarking, not clinical or unattended deployment. Strict online operation and end-to-end notification latency remain unmeasured.