Phase-Aware Spatio-Temporal Deep Learning for Contextual Fall Detection in Geriatric Environments

Abstract

In open living spaces, traditional vision-based fall detection architectures have been found to have high False-Alarm Rates (FAR) because they cannot differentiate between real falls and high velocity Activities of Daily Living (ADLs) with similar footprints of deformation in space. This paper proposes the basic subsystem of an overall multi-zone ambient intelligence system for automated geriatric care. In this study, focusing on open room environment monitoring, a hybrid Spatio-Temporal Deep Learning model is proposed where the human fall is considered as a deterministic sequence of three chronological stages: pre-impact kinematics, critical fall, and post-fall immobility. This system uses a 3D Convolutional Neural Network (3D-CNN) spatial feature extractor and a sequence-mapping Temporal Convolutional Network (TCN) to continuously assess spatial volumes as well as real-time bounding box aspect ratios and vertical velocity vectors. Experimental testing on public UR Fall Detection (URFD) and UP-Fall Detection datasets shows that this phase-aware approach can effectively reduce the number of false alarms that are triggered by fast actions, with a targeted reduction of 15% and an overall classification accuracy of 99.12% and a high detection sensitivity of 99.32%. The connection of this kinematic foundation with future non-invasive pain analytics and privacy-preserving bathroom pipelines based on non-RGB is described.