Real-Time Fall Detection in Indoor Environments Using FMCW Radar and Lightweight Ensemble CNN on Edge Devices.

Falls are a leading cause of injury and mortality among the elderly, particularly indoors where privacy concerns limit real-time monitoring. This study developed a privacy-preserving, real-time fall detection system using a frequency-modulated continuous wave radar sensor and a lightweight convolutional neural network (CNN) architecture deployed on an edge device. To address the limitations of wearable and vision-based methods, the system leverages contactless radar sensing and an ensemble of two independent CNN models, namely, a base model designed to capture the entire fall sequence for high recall and a sub-model aimed at pinpointing the fall moment with high precision. The final decision is executed through a voting mechanism that triggers a fall alert only when both models agree, significantly reducing false positives. On 14 recording sessions withheld from the training of both models, the ensemble detected 13 of 14 falls (recall 92.9%, 95% CI 68.5-98.7%) without false alarms, and in a 24-trial on-device test comprising 4 falls and 20 non-fall activities it detected every fall with no false alarm. Running on a low-power microcontroller, the system indicates the feasibility of privacy-preserving on-device fall detection, while validation with older adults and in real living environments remains necessary.