Abstract
<title>Abstract</title> <p>Rapid and highly specific fall detection is critical to preserving the health and autonomy of vulnerable elderly populations. While existing vision-based fall detection architectures promise non-intrusive monitoring, they suffer from high false positive rates triggered by ordinary Activities of Daily Living (ADLs) and severe performance degradation under joint occlusions. This paper presents DAÏO (Intelligent Vision-Based Fall Detection System), a novel, edge-capable, high-specificity computer vision framework designed to address these limitations. DAÏO operates on a hybrid architecture combining a lightweight human pose estimator, an Extended Kalman Filter (EKF) with adaptive measurement noise, and a stateful Triple-Check Logic Engine. By tracking joint dynamics and mapping them through an adaptive mathematical state-space model, the system adjusts to occlusion events dynamically using visibility confidence scores. Falling actions are validated across spatial, rotational, and kinematic dimensions in parallel, filtering out high-velocity non-fall events (such as prostrations, deep squats, and rapid sitting) via a velocity latching mechanism. Evaluated against a composite dataset comprising 340 video sequences (~153,000 frames) from the UR Fall Dataset, the Multiple Cameras Fall Dataset (MCFD), and customized stress-test scenarios, the system achieved an overall accuracy of 95.3%, a sensitivity of 93.3%, and a specificity of 96.4%. Crucially, the velocity latch cut false positive alarms by 79.7% compared to baseline posture-only models, demonstrating 100% specificity during active, repetitive religious prostrations. Running locally at 35.7 FPS on consumer workstations and 12.3% on a Raspberry Pi 4, the proposed system demonstrates high specificity, operational reliability, and privacy preservation at the network edge.</p>