Human Fall Detection Through Velocity-Driven Temporal Deep Learning

Accidental falls remain one of the most frequent and dangerous incidents among older adults, often leading to hospitalization, long-term disability, and loss of independence. This research presents a novel, non-wearable, vision-based system that addresses the challenge of real-time fall detection and instability recognition in home environments. The system uses monocular-camera-captured video to track human motion and classify it into three categories: normal, fall, or unstable. By employing YOLOv8 for keypoint extraction and a two-stage deep learning framework, the model predicts human movement velocity and classifies activities based on temporal skeleton keypoint dynamics. The two-stage framework consists of two separate LSTM-based models: the first stage predicts velocity from keypoint sequences using regression, and the second stage classifies motion into three categories. Evaluation on a primary dataset of 3449 annotated high-resolution video clips demonstrates a 90% classification accuracy. Compared to baseline methods, the proposed system improves the accuracy of fall detection, showing strong potential for practical deployment in elderly care settings. The approach offers a scalable and unobtrusive solution for continuous monitoring, eliminating the need for wearable devices.