Explainable Domain-Robust Fall Detection Using Temporal Pose Features

Abstract

Vision-based fall detection can fail when camera geometry and acquisition conditions change. This study developed an explainable temporal pose pipeline and evaluated it using progressively stricter leakage-safe protocols. YOLO11n-Pose key points were transformed into robust temporal descriptors. Video-independent URFD validation achieved 81.67% balanced accuracy, whereas unchanged transfer to MCFD produced 46.99%, confirming substantial domain shift. Under simultaneous unseen-camera and unseen-scenario testing, invariant-feature learning achieved 62.21%; offline complete-interval summarization increased this upper bound to 71.16%. In a distinct fixed-installation experiment, camera-aware synchronized multi-view fusion achieved 87.23% balanced accuracy, 82.35% sensitivity and 92.11% specificity on unseen chute scenarios. Its improvement over permutation-invariant fusion was not statistically significant. The findings show that temporal evidence and synchronized views can improve recognition under specified deployment conditions, while fixed-camera multi-view performance must not be interpreted as unseen-camera generalization or real-time clinical readiness.