| Title |
Tiny ML and Edge Computing-based Occupant Activity Diagnosis Technology |
| Authors |
김지형(Kim, Ji-Hyeoung) ; 김선인(Kim, Seon-In) ; 박영준(Park, Young-Joon) ; 김의종(Kim, Eui-Jong) |
| DOI |
https://doi.org/10.5659/JAIK.2026.42.3.283 |
| Keywords |
RADAR sensor; Fall detection; Tiny machine learning |
| Abstract |
This study presents a real-time indoor fall detection method that leverages Tiny ML and edge computing to address the network dependency
and privacy concerns inherent in traditional cloud-based monitoring systems. Fall motion labeling was performed on time-series data collected
from an FMCW radar sensor. The signals were preprocessed using sliding window segmentation, Hamming smoothing, and Fourier transform
to reduce noise and extract features suitable for training. A lightweight 1D CNN?LSTM hybrid model was developed and trained. To
overcome the limitations of traditional time-based evaluation?where intermittent predictions can distort true performance?we implemented an
event-based evaluation and correction procedure. This method detects the start, peak, and end of events and merges adjacent predictions into
single episodes. Our approach achieved a recall of 94.7% and an F1-score of 83.7% on the test dataset. Furthermore, the model was
successfully deployed on a resource-constrained microcontroller unit (MCU), demonstrating an inference time of 15 ms and efficient memory
usage (136 KB Flash and 520 KB RAM), which confirms its feasibility for real-time edge processing. Limitations include data collected from
a single sensor in a controlled environment and the absence of on-device embedding evaluation. Future work will expand to include a
broader range of human activities, enhance dataset diversity across multiple environments and subjects, and conduct long-term field testing in
real-world residential settings to improve generalization and reliability. |