Vision-Based Fall Detection in Hospitals Using a Dual-YOLO Deep Learning Framework

Abstract

Each year, an estimated 700,000 to 1,000,000 patients experience falls in hospitals across the United States [1], and inpatient falls are similarly rising in Indian healthcare settings [25]. Manual CCTV surveillance remains the norm but cannot reliably cover multiple feeds, especially during understaffed night shifts, leaving many preventable falls undetected in time. This paper’s central novelty is coupling a dual-layer detection engine; a geometric aspect-ratio screener (YOLOv10s) running in parallel with a learned fall classifier (YOLOv11n); with multi-frame temporal validation, rather than relying on either mechanism alone. Layer 1 flags a width-to-height ratio exceeding one as a geometric sign of being prone; Layer 2, trained on 8,749 indoor fall images, provides semantic confirmation. A fall is escalated only if both layers agree for 4 consecutive frames. ByteTrack preserves identity continuity through brief occlusions, and a 10×10 spatial grid filter suppresses false alarms from static furniture. On the held-out validation set the system reaches 92.43% precision, 89.05% recall, and 95.17% mAP@50 at 30 fps on a single consumer GPU, with the four-frame confirmation alone cutting false alerts by 71.1% relative to single-frame detection. Confirmed falls trigger simultaneous email, SMS, and WhatsApp alerts within ≈4.2 s, plus a Flask dashboard for live monitoring. Unlike transformer-based alternatives, the system runs entirely on commodity single-GPU hardware, underscoring real-world deployability; its main limitations ; single-camera operation, degraded low-light performance, and staged (non-clinical) training data ; are discussed in Section VI.